{"schemaVersion":1,"language":"en","categories":["Fundamentals","Retrieval & Context","Agents & Interfaces","Training & Adaptation","Inference & Operations","Quality & Governance","Modality & Production"],"topics":[{"id":"llm-fundamentals","number":1,"title":"LLM Fundamentals","category":"Fundamentals","summary":"What a large language model learns, generates, and does not guarantee.","definition":"A large language model estimates probabilities for linguistic continuations from learned parameters. It produces plausible sequences but has, on its own, neither current source authority nor a proof of truth for individual statements.","importance":"Without this basic picture, fluency, knowledge, reasoning, and verification are easily confused with one another.","boundary":"An output that sounds probable is neither automatically true nor traceable to a specific source.","keywords":["language model","probability","parameter","next token","generative"],"related":["transformers","tokenization","llm-inference","hallucination-detection"]},{"id":"transformers","number":2,"title":"Transformers","category":"Fundamentals","summary":"An architecture that processes sequences through attention instead of recurrent state chains.","definition":"Transformers consist of repeated attention and feed-forward blocks. They model relationships between positions in an input sequence and can be trained with a high degree of parallelism.","importance":"They form the technical foundation of most modern language models and of many multimodal models.","boundary":"The architecture says nothing about whether a result is correct, current, or authorized.","keywords":["encoder","decoder","self attention","feed forward","sequence"],"related":["attention","multi-head-attention","transformer-block","residual-stream","positional-encoding","llm-fundamentals"]},{"id":"attention","number":3,"title":"Attention Mechanism","category":"Fundamentals","summary":"Each position mixes the positions it may read, with weights that add up to 100%.","plain":"Each word looks at the earlier words and decides how much each of them matters to it.","definition":"For a query position, attention scores every key it may read, turns the scores into weights with softmax and adds up the values with those weights. Each query head computes its own weights; in GPT-2-style models each head also has its own key and value projections, while with grouped-query attention (Llama, Qwen2) several query heads share one key/value head.","importance":"This lets models capture long-range references and context-dependent relationships.","boundary":"Attention weights are not a reliable explanation and not a proof of causal use.","keywords":["query","key","value","weights","self attention"],"related":["transformers","multi-head-attention","attention-head","causal-mask","softmax","embeddings","observability"]},{"id":"multi-head-attention","number":4,"title":"Multi-Head Attention","category":"Fundamentals","summary":"Several attention projections view the same sequence in parallel.","definition":"Multi-head attention splits the representation into several learned attention heads. Each head can capture different relationships or projections; the results are then merged.","importance":"The parallel views increase the expressiveness of a transformer block.","boundary":"A head has no firmly guaranteed semantic role unless that role has been measured separately.","keywords":["attention heads","projection","parallel","head anatomy"],"related":["attention","attention-head","transformer-block","transformers","llm-evaluation"]},{"id":"tokenization","number":5,"title":"Tokenization","category":"Fundamentals","summary":"The breaking of text into discrete model units.","plain":"Cutting text into small pieces (tokens): whole words, parts of words or single characters.","definition":"Tokenization translates text into IDs a model can process. Depending on the method, tokens correspond to words, word parts, characters, or bytes.","importance":"It affects sequence length, multilingualism, number representation, and the model's view of spelling variants.","boundary":"Token boundaries are technical boundaries and not automatically linguistic or ontological ones.","keywords":["bpe","sentencepiece","byte","subword","vocabulary"],"matchTerms":["tokenizer","tokenizers"],"related":["embeddings","context-windows","llm-fundamentals"]},{"id":"embeddings","number":6,"title":"Embeddings","category":"Fundamentals","summary":"Learned vectors for discrete or structured objects.","definition":"Embeddings map tokens, sentences, documents, or other objects to dense numeric vectors. Similar vectors can indicate statistically similar usage.","importance":"They make discrete inputs processable for neural models and power semantic search methods.","boundary":"Vector proximity is not identity, not truth, and not an authority relationship.","keywords":["vector","representation","similarity","dense"],"related":["token-embedding","semantic-search","vector-databases","attention"]},{"id":"positional-encoding","number":7,"title":"Positional Encoding","category":"Fundamentals","summary":"Information about where an element sits in a sequence.","plain":"Extra information that tells the model at which place in the text each piece stands.","definition":"Because attention alone knows no order, models add position information: GPT-2-style models add a learned position vector to each token embedding, rotary encoding (RoPE, used by Llama and Qwen2) rotates queries and keys by position inside attention, and relative schemes bias the attention scores. This lets the model distinguish order and distance.","importance":"The choice affects length generalization and the learnability of relational order.","boundary":"A position signal can create unwanted shortcuts when position and target correlate during training.","keywords":["rope","absolute position","relative position","sequence order"],"related":["transformers","attention","context-windows"]},{"id":"prompt-engineering","number":8,"title":"Prompt Engineering","category":"Fundamentals","summary":"Systematic design of the instruction and the provided context.","definition":"Prompt engineering structures task, examples, output format, boundaries, and context so that a model responds more reliably in a desired role.","importance":"It is a fast lever for behavior, especially without training.","boundary":"A good prompt is neither a security mechanism nor a guarantee of factual or tool fidelity.","keywords":["system prompt","instruction","template","prompt design"],"related":["few-shot-learning","structured-outputs","prompt-evaluation","guardrails"]},{"id":"few-shot-learning","number":9,"title":"Few-Shot Learning","category":"Fundamentals","summary":"Task steering through a few examples in the context.","definition":"In few-shot learning the model receives a few input-output examples inside the prompt and infers the desired pattern for new cases from them.","importance":"It can convey format and task schema without a weight update.","boundary":"Examples consume context and can anchor unwanted patterns or errors as a template.","keywords":["in-context learning","examples","demonstrations","zero shot"],"related":["prompt-engineering","context-windows","fine-tuning"]},{"id":"chain-of-thought","number":10,"title":"Chain-of-Thought","category":"Fundamentals","summary":"Intermediate steps a model produces before or alongside its answer.","definition":"Chain-of-thought refers to textual intermediate reasoning that can break complex tasks into smaller steps. Visible reasoning can be helpful, but is itself model text.","importance":"The technique can improve handling of multi-step tasks and make failure points visible.","boundary":"A plausible chain of thought is no proof that exactly those steps were internally causal or factually correct.","keywords":["reasoning trace","scratchpad","reasoning","intermediate steps"],"related":["structured-outputs","llm-evaluation","observability"]},{"id":"rag","number":11,"title":"RAG","category":"Retrieval & Context","summary":"Retrieval-augmented generation combines search with text generation.","definition":"RAG retrieves relevant content at runtime and gives it to a generator as context. This means current or private facts do not have to reside entirely in model weights.","importance":"It improves updatability and can enable source attribution.","boundary":"Found does not mean admitted; context does not mean evidence; citing does not mean the statement follows from the source.","keywords":["retrieval augmented generation","grounding","documents","context"],"related":["semantic-search","reranking","chunking","vector-databases"]},{"id":"vector-databases","number":12,"title":"Vector Databases","category":"Retrieval & Context","summary":"Storage and indexes for similarity search over vectors.","definition":"A vector database stores embeddings together with IDs and metadata and finds entries by a distance or similarity measure.","importance":"It makes semantic search over large collections practical.","boundary":"The index decides proximity, not validity, currency, rights, or contradiction.","keywords":["ann","index","cosine similarity","embedding store"],"related":["embeddings","semantic-search","rag","hybrid-search"]},{"id":"semantic-search","number":13,"title":"Semantic Search","category":"Retrieval & Context","summary":"Search by proximity of meaning rather than only identical character strings.","definition":"Semantic search encodes query and content into representations and ranks candidates by modeled proximity of meaning.","importance":"It finds paraphrased or conceptually related content that a pure keyword search misses.","boundary":"Semantic proximity can confidently deliver wrong candidates and needs open-set and authority boundaries.","keywords":["dense retrieval","similarity","meaning","nearest neighbor"],"related":["embeddings","vector-databases","hybrid-search","reranking"]},{"id":"hybrid-search","number":14,"title":"Hybrid Search","category":"Retrieval & Context","summary":"A combination of lexical and semantic search.","definition":"Hybrid search combines exact or term-based hits with embedding-based similarity and merges both result lists through rules or ranking.","importance":"It covers both precise names and IDs as well as linguistic variants.","boundary":"The fusion needs a defined calibration; two scores from different spaces are not automatically comparable.","keywords":["bm25","dense sparse","fusion","lexical search"],"related":["semantic-search","reranking","vector-databases"]},{"id":"reranking","number":15,"title":"Reranking","category":"Retrieval & Context","summary":"A second, more precise sorting step for candidates that were already found.","definition":"A reranker re-scores a small candidate set with more context or a stronger model. It optimizes the ordering, not the original coverage.","importance":"It can place relevant evidence earlier and limit the active payload.","boundary":"A reranker cannot bring back a missing candidate and must not confuse ranking with admission.","keywords":["cross encoder","ranking","candidate set","relevance"],"related":["semantic-search","hybrid-search","rag","model-routing"]},{"id":"chunking","number":16,"title":"Chunking Strategies","category":"Retrieval & Context","summary":"How sources are split into retrievable sections.","definition":"Chunking divides documents into retrieval units by length, sentences, sections, structure, or semantic boundaries.","importance":"The split determines whether enough context is included without blurring the search with large blocks.","boundary":"A chunk must not lose its source spans, version, and neighborhood; otherwise later provenance is only approximate.","keywords":["document splitting","span","overlap","segmentation"],"related":["rag","context-windows","semantic-search"]},{"id":"context-windows","number":17,"title":"Context Windows","category":"Retrieval & Context","summary":"The maximum active input sequence of a model call.","definition":"The context window includes system instruction, user input, examples, retrieved content, and prior output that a model can take into account in one call.","importance":"It limits active information, cost, and runtime and influences which content is processed together.","boundary":"More context guarantees no better use; irrelevant or contradictory content can degrade performance.","keywords":["sequence length","tokens","long context","prompt size"],"related":["context","chunking","rag","kv-cache","cost-optimization"]},{"id":"function-calling","number":18,"title":"Function Calling","category":"Agents & Interfaces","summary":"Model output in a predefined function schema.","definition":"Function calling lets a model propose a function name and structured arguments instead of only producing free text.","importance":"It creates a machine-readable seam between language and program code.","boundary":"Schema-conformant arguments are not automatically sensible, safe, or authorized.","keywords":["json schema","arguments","api call","structured call"],"related":["tool-calling","structured-outputs","mcp"]},{"id":"tool-calling","number":19,"title":"Tool Calling","category":"Agents & Interfaces","summary":"A model proposes the use of an external capability.","definition":"Tool calling extends function calling with concrete external actions such as search, database access, code execution, or business operations.","importance":"This lets a language model obtain information or affect real systems.","boundary":"The model proposal must never alone be the execution authorization; errors, partial results, and side effects need their own contracts.","keywords":["tools","actions","execution","transaction"],"related":["function-calling","ai-agents","mcp","guardrails"]},{"id":"ai-agents","number":20,"title":"AI Agents","category":"Agents & Interfaces","summary":"Systems that pursue goals across several model and tool steps.","definition":"An AI agent combines model decisions, state, tools, and control logic to work on a task over multiple steps.","importance":"Agents can solve dynamic tasks where the next step only emerges from the previous result.","boundary":"More autonomy enlarges error chains and side effects; a loop is not governance.","keywords":["agent","autonomy","planning","loop"],"related":["agentic-workflows","tool-calling","memory-systems","guardrails"]},{"id":"agentic-workflows","number":21,"title":"Agentic Workflows","category":"Agents & Interfaces","summary":"Explicit flow structure for agentic steps and decisions.","definition":"Agentic workflows arrange model calls, tools, checks, retries, and handoffs into a controlled process.","importance":"A workflow makes autonomy boundable and allows specialized components instead of an unclear endless loop.","boundary":"A graphical flow proves no correct authority or error handling; every edge needs semantics.","keywords":["workflow","orchestration","state machine","loop"],"related":["ai-agents","multi-agent-systems","observability","production-genai"]},{"id":"memory-systems","number":22,"title":"Memory Systems","category":"Agents & Interfaces","summary":"Persistent or runtime-adjacent state outside the model context.","definition":"Memory systems store conversations, observations, facts, working state, or summaries beyond individual model calls.","importance":"They enable continuity and reduce the need to put everything into the prompt again.","boundary":"Stored does not mean true; memories need provenance, version, validity, and deletion or revocation rules.","keywords":["long term memory","state","store","conversation memory"],"related":["rag","ai-agents","vector-databases","observability"]},{"id":"multi-agent-systems","number":23,"title":"Multi-Agent Systems","category":"Agents & Interfaces","summary":"Several separate agent roles work on the same task.","definition":"Multi-agent systems distribute planning, research, critique, or execution across several model- or rule-based actors with defined message paths.","importance":"Specialization and cross-checking can structure complex work.","boundary":"Agreement among several models is not independent truth; shared training imprint can multiply the same error.","keywords":["agents","debate","orchestration","roles"],"related":["ai-agents","agentic-workflows","llm-evaluation"]},{"id":"mcp","number":24,"title":"MCP","category":"Agents & Interfaces","summary":"An open protocol for tools, resources, and prompts between AI hosts and servers.","definition":"The Model Context Protocol standardizes communication between host, clients, and specialized servers. Servers can offer clearly declared resources, tools, and prompts; capabilities are negotiated explicitly.","importance":"It reduces proprietary point-to-point integrations and makes specialized services combinable.","boundary":"MCP standardizes the transport and the capability surface, not truth, tool authorization, or evidence quality.","keywords":["model context protocol","host client server","resources","tools","prompts"],"related":["tool-calling","function-calling","ai-agents"],"reference":"https://modelcontextprotocol.io/docs/learn/architecture"},{"id":"structured-outputs","number":25,"title":"Structured Outputs","category":"Agents & Interfaces","summary":"Model outputs that follow a fixed machine-readable schema.","definition":"Structured outputs constrain the output to fields, types, and allowed values, usually via JSON Schema, grammar, or constrained decoding.","importance":"They make outputs validatable and reduce parsing ambiguity.","boundary":"Structurally valid is not semantically correct; a perfectly valid object can carry wrong values.","keywords":["json","schema","grammar constrained","typed output"],"related":["function-calling","guardrails","llm-evaluation"]},{"id":"optimizer","number":26,"title":"Optimizer","category":"Training & Adaptation","summary":"The rule that turns gradients into parameter updates during training.","definition":"An optimizer uses gradients and a learning rate to update model parameters. Methods differ in how they combine past gradients, scale individual updates, and apply weight decay.","importance":"It affects training stability, memory use, and how quickly a model reaches a useful loss.","boundary":"An optimizer does not define the training objective or guarantee generalization. This app currently uses AdamW rather than offering an optimizer selector.","keywords":["gradient descent","learning rate","parameter update","optimizer state"],"related":["sgd","rmsprop","adam","adamw","fine-tuning"]},{"id":"sgd","number":27,"title":"SGD","category":"Training & Adaptation","summary":"Stochastic gradient descent updates weights from sampled-batch gradients.","definition":"Plain SGD subtracts the learning-rate-scaled gradient from each parameter. Momentum is an optional variant that carries part of earlier updates forward.","importance":"Its simple update makes optimizer behavior and the effect of learning rate easy to inspect.","boundary":"Plain SGD has little persistent state, but can need careful learning-rate tuning. It is a comparison concept, not a selectable optimizer in this app.","keywords":["stochastic gradient descent","momentum","learning rate","minibatch","optimizer"],"related":["optimizer","rmsprop","adam","adamw"],"reference":"https://docs.pytorch.org/docs/stable/generated/torch.optim.SGD.html"},{"id":"rmsprop","number":28,"title":"RMSprop","category":"Training & Adaptation","summary":"An adaptive update scaled by a running average of squared gradients.","definition":"RMSprop tracks recent squared gradients per parameter and divides the gradient by their root-scale estimate, with a small stabilizing term.","importance":"It illustrates why different parameters can receive different effective step sizes.","boundary":"Adaptive steps do not guarantee better validation loss or prevent overfitting. RMSprop is not selectable in this app.","keywords":["adaptive learning rate","squared gradient","moving average","optimizer"],"related":["optimizer","sgd","adam","adamw"],"reference":"https://docs.pytorch.org/docs/stable/generated/torch.optim.RMSprop.html"},{"id":"adam","number":29,"title":"Adam","category":"Training & Adaptation","summary":"Adaptive Moment Estimation combines moving averages of gradients and their squares.","definition":"Adam keeps a first-moment estimate of gradients and a second-moment estimate of squared gradients for each parameter, then uses them to scale updates.","importance":"It explains the extra optimizer-state memory shown when planning a run.","boundary":"Adam uses additional state and does not itself ensure generalization. The app uses AdamW, not a selectable Adam option.","keywords":["adaptive moment estimation","first moment","second moment","optimizer state"],"related":["optimizer","rmsprop","adamw"],"reference":"https://docs.pytorch.org/docs/stable/generated/torch.optim.Adam.html"},{"id":"adamw","number":30,"title":"AdamW","category":"Training & Adaptation","summary":"Adam-style updates with weight decay separated from the gradient step.","definition":"AdamW retains Adam's first and second moment estimates but applies weight decay separately from the adaptive gradient update. This app uses AdamW for pre-training and fine-tuning.","importance":"The separation makes the role of weight decay clearer, while the two moment buffers remain part of the memory plan.","boundary":"The app does not offer an optimizer choice; AdamW and weight decay still require validation rather than guaranteeing better outputs.","keywords":["decoupled weight decay","adam","optimizer state","pretraining","sft"],"related":["optimizer","adam","fine-tuning","sft"],"reference":"https://docs.pytorch.org/docs/stable/generated/torch.optim.AdamW.html"},{"id":"fine-tuning","number":31,"title":"Fine-Tuning","category":"Training & Adaptation","summary":"Further training of a pretrained model for a task or domain.","definition":"In fine-tuning, model parameters are updated with a smaller, targeted dataset. This can specialize behavior, format, or domain performance.","importance":"It shifts recurring patterns out of the prompt and into the weights.","boundary":"Fine-tuning guarantees neither generalization nor current facts and can shift existing capabilities.","keywords":["supervised fine tuning","sft","training","weights"],"related":["sft","lora-qlora","peft","dpo","rlhf"]},{"id":"sft","number":32,"title":"SFT","category":"Training & Adaptation","summary":"Supervised fine-tuning on demonstrated assistant responses.","definition":"SFT continues training a pretrained model on example conversations. In this app the full conversation supplies context, but only assistant response tokens contribute to the training loss.","importance":"It teaches a base model to answer in the demonstrated format instead of merely continuing raw text.","boundary":"SFT learns from the examples it receives; it does not prove factual accuracy, preference alignment, or safety. This app updates model weights rather than training LoRA adapters.","keywords":["supervised fine tuning","instruction tuning","assistant loss mask","conversation examples"],"related":["fine-tuning","dpo","peft","adamw"]},{"id":"lora-qlora","number":33,"title":"LoRA & QLoRA","category":"Training & Adaptation","summary":"Low-parameter adaptation via low-rank add-on matrices.","definition":"LoRA freezes the base matrix and trains small low-rank adapters. QLoRA combines this principle with a quantized base model to reduce memory use further.","importance":"This makes it possible to adapt large models with far fewer trainable parameters.","boundary":"Fewer trainable parameters do not automatically mean less data need, better quality, or easier evaluation. LoRA and QLoRA are not offered in this app, which trains all weights of its own GPT checkpoints.","keywords":["low rank","adapter","quantized training","qlora"],"related":["peft","fine-tuning","quantization"]},{"id":"peft","number":34,"title":"PEFT","category":"Training & Adaptation","summary":"A family of parameter-efficient fine-tuning methods.","definition":"Parameter-efficient fine-tuning changes only a small part of additional or selected parameters, such as adapters, prefixes, or low-rank matrices.","importance":"It lowers memory and compute cost for many specialized variants.","boundary":"PEFT is a training method, not a guarantee of isolation between roles or of safe adapter combination. This app offers no PEFT method; pretraining, SFT, DPO and ORPO update all weights of app-trained GPT checkpoints.","keywords":["parameter efficient","adapters","prefix tuning","lora"],"related":["lora-qlora","fine-tuning","model-compression"]},{"id":"rlhf","number":35,"title":"RLHF","category":"Training & Adaptation","summary":"Optimizing a model with feedback learned from human preferences.","definition":"Reinforcement Learning from Human Feedback uses human comparisons, a preference or reward model, and subsequent optimization to align responses with desired behavior.","importance":"It can improve helpfulness, style, and safety behavior.","boundary":"Preference is not truth; reward hacking and unclear reviewer standards remain possible. RLHF with a reward model is not implemented in this app; its DPO and ORPO modes learn from offline preference pairs instead.","keywords":["reward model","human feedback","preference","reinforcement learning"],"related":["dpo","fine-tuning","ai-safety","llm-evaluation"]},{"id":"dpo","number":36,"title":"DPO","category":"Training & Adaptation","summary":"Direct optimization toward preferred over rejected responses.","definition":"Direct Preference Optimization trains a model directly from response pairs, without running a separate classic RL loop with its own reward model.","importance":"It simplifies preference alignment compared to many RLHF pipelines.","boundary":"The method inherits the quality, bias, and ontology of the preference data and proves no factual fidelity.","keywords":["direct preference optimization","chosen rejected","preference data"],"related":["sft","rlhf","fine-tuning","prompt-evaluation"]},{"id":"knowledge-distillation","number":37,"title":"Knowledge Distillation","category":"Training & Adaptation","summary":"A smaller model learns from the outputs or representations of a larger teacher.","definition":"Distillation transfers a teacher model's behavior to a smaller student, for example via soft targets, logits, intermediate states, or generated training examples.","importance":"It can produce compact specialists that run locally and cheaply.","boundary":"The student also inherits the teacher's errors and blind spots; teacher agreement is no independent gold standard. Distillation is not a training mode in this app.","keywords":["teacher student","soft targets","logits","student model"],"related":["model-compression","fine-tuning","llm-evaluation"]},{"id":"quantization","number":38,"title":"Quantization","category":"Training & Adaptation","summary":"Representing weights or activations with fewer bits.","definition":"Quantization reduces numeric precision, for example from FP32 to FP16, INT8, or lower formats. It can happen after training or during quantization-aware training.","importance":"It lowers memory use and can speed up inference.","boundary":"Rounding and scaling errors can create rare functional failures that an aggregate hides.","keywords":["int8","int4","fp16","precision","rounding"],"related":["model-compression","llm-inference","lora-qlora"]},{"id":"model-compression","number":39,"title":"Model Compression","category":"Training & Adaptation","summary":"Methods to reduce model size or compute cost.","definition":"Model compression covers quantization, pruning, distillation, factorization, and other methods that make a model smaller or faster.","importance":"It enables local execution and lower operating resources.","boundary":"Compression is only a success if the relevant quality and safety axes are preserved.","keywords":["pruning","distillation","quantization","size reduction"],"related":["quantization","knowledge-distillation","cost-optimization"]},{"id":"llm-inference","number":40,"title":"LLM Inference","category":"Inference & Operations","summary":"Running a trained model on new inputs.","definition":"Inference covers preprocessing, forward passes, decoding, cache management, and output preparation of an already trained model.","importance":"This is where real latency, memory use, parallelism, and runtime behavior arise.","boundary":"A correctly loaded checkpoint and deterministic decoding prove no factual correctness.","keywords":["forward pass","decode","runtime","generation"],"related":["forward-pass","sampling","kv-cache","speculative-decoding","llm-serving","quantization"]},{"id":"kv-cache","number":41,"title":"KV Cache","category":"Inference & Operations","summary":"A cache for attention keys and values of already processed tokens.","definition":"In autoregressive decoding, the KV cache stores computed key and value states so that earlier tokens are not fully reprocessed for each new token.","importance":"It speeds up generation greatly but can occupy a lot of memory.","boundary":"The cache stores keys and values, not attention weights or outputs: each new query still computes fresh weights over all cached keys. With grouped-query attention it holds one entry per key/value head, not per query head. Decoding with a cache and recomputing the whole context are the same calculation on paper but not bit for bit; in bfloat16 the logits can differ noticeably.","keywords":["key value cache","decode","memory","autoregressive"],"related":["llm-inference","context-windows","vllm"]},{"id":"speculative-decoding","number":42,"title":"Speculative Decoding","category":"Inference & Operations","summary":"A fast draft is verified in bulk by a target model.","definition":"In speculative decoding, a smaller or cheaper method produces several token proposals that the target model accepts or rejects in fewer steps.","importance":"It can lower latency without deliberately changing the target distribution.","boundary":"The method speeds up decoding but solves no evidence, hallucination, or authority question.","keywords":["draft model","verification","decode acceleration","tokens"],"related":["llm-inference","llm-serving","vllm"]},{"id":"llm-serving","number":43,"title":"LLM Serving","category":"Inference & Operations","summary":"Providing models as robust, plannable runtime services.","definition":"LLM serving covers model loading, request queue, batching, parallelism, timeouts, streaming, scaling, and error isolation.","importance":"It decides whether a model is reproducible and operable under real load.","boundary":"High throughput is no proof of quality or governance; serving must not bypass domain gates.","keywords":["deployment","batching","throughput","latency","runtime"],"related":["vllm","llm-inference","observability","production-genai"]},{"id":"vllm","number":44,"title":"vLLM","category":"Inference & Operations","summary":"A specialized open-source serving stack for high LLM load.","definition":"vLLM is an inference and serving engine with memory management for the KV cache, continuous batching, and further optimized execution paths.","importance":"It can provide high throughput for compatible models and hardware.","boundary":"vLLM is infrastructure: it defines neither evidence lineage nor model role, tool authority, or claim validation.","keywords":["pagedattention","continuous batching","serving engine","kv cache"],"related":["llm-serving","kv-cache","speculative-decoding"],"reference":"https://docs.vllm.ai/en/latest/"},{"id":"model-routing","number":45,"title":"Model Routing","category":"Inference & Operations","summary":"A request is assigned to a suitable model or path.","definition":"Model routing selects between several models and processing paths based on task, domain, risk, cost, or availability.","importance":"It allows specialization and prevents every request from using the most expensive monolith.","boundary":"A wrong router can produce internally plausible results in the wrong scope; unknown must be its own exit.","keywords":["router","scope selector","mixture of experts","fallback"],"related":["ai-agents","reranking","production-genai"]},{"id":"guardrails","number":46,"title":"Guardrails","category":"Quality & Governance","summary":"Technical boundaries for inputs, outputs, and actions.","definition":"Guardrails check or limit formats, content, values, permissions, and actions before, during, or after a model call.","importance":"They reduce clearly named error classes and can enforce fail-closed behavior.","boundary":"A downstream filter cannot retroactively invent missing evidence and must not become the hidden decider.","keywords":["validation","policy","filter","fail closed"],"related":["ai-safety","structured-outputs","hallucination-detection","tool-calling"]},{"id":"ai-safety","number":47,"title":"AI Safety","category":"Quality & Governance","summary":"Recognizing, limiting, and demonstrating risks of AI systems.","definition":"AI safety covers technical and organizational methods against misbehavior, misuse, loss of control, unwanted side effects, and unsustainable decisions.","importance":"Capability without limited consequences is not a sufficient goal in real systems.","boundary":"Safety is not a single metric and not a prompt; it depends on system context, rights, and error costs.","keywords":["risk","alignment","control","misuse","robustness"],"related":["guardrails","observability","llm-evaluation","production-genai"]},{"id":"hallucination-detection","number":48,"title":"Hallucination Detection","category":"Quality & Governance","summary":"Checking whether a generated statement is unsupported or contradictory.","definition":"Hallucination detection compares outputs with sources, rules, other models, or consistency signals to flag unsupported statements.","importance":"It can make errors visible before delivery.","boundary":"A detector is itself error-prone; after-the-fact detection replaces no architecture that binds claims to evidence.","keywords":["factuality","unsupported claim","verification","grounding"],"related":["rag","guardrails","llm-evaluation","structured-outputs"]},{"id":"llm-evaluation","number":49,"title":"LLM Evaluation","category":"Quality & Governance","summary":"Systematic measurement of model behavior on defined surfaces.","definition":"LLM evaluation combines datasets, metrics, baselines, human judgments, counterfactuals, and robustness tests for clearly delimited questions.","importance":"Only a bound evaluation separates progress from impression, prompt luck, or test fitting.","boundary":"An aggregate metric can hide functional failures, seed tails, coverage, and abstention.","keywords":["benchmark","metrics","gold","baseline","counterfactual"],"related":["prompt-evaluation","observability","ai-safety"]},{"id":"prompt-evaluation","number":50,"title":"Prompt Evaluation","category":"Quality & Governance","summary":"Comparing prompt variants on a fixed task and sample.","definition":"Prompt evaluation measures how instructions, examples, and format requirements affect defined outputs.","importance":"It prevents individual attractive examples from counting as a general improvement.","boundary":"Prompt quality is only one component; provider, model, data, and decoder changes must be separated.","keywords":["prompt testing","ab test","template","provider"],"related":["prompt-engineering","llm-evaluation","few-shot-learning"]},{"id":"observability","number":51,"title":"Observability","category":"Quality & Governance","summary":"Understanding from runtime data what a system did and where it failed.","definition":"Observability connects structured logs, traces, metrics, and artifact versions across components and requests.","importance":"It makes errors, cost, latency, and drift investigable without guessing the run after the fact.","boundary":"A log shows an event, but not automatically its domain justification or causal meaning.","keywords":["logs","traces","metrics","telemetry","lineage"],"related":["production-genai","llm-evaluation","ai-safety"]},{"id":"cost-optimization","number":52,"title":"AI Cost Optimization","category":"Quality & Governance","summary":"Quality-bound reduction of compute, memory, and operating cost.","definition":"Cost optimization uses routing, smaller models, batching, caching, quantization, limited payloads, and better utilization.","importance":"It decides whether a system is economically scalable.","boundary":"A smaller or cheaper output is no success if evidence coverage, tail safety, or task quality drops.","keywords":["latency","tokens","gpu utilization","payload","economics"],"related":["model-routing","quantization","model-compression","llm-serving"]},{"id":"multimodal-ai","number":53,"title":"Multimodal AI","category":"Modality & Production","summary":"Joint processing of several modalities such as text, image, and audio.","definition":"Multimodal systems connect representations and tasks across text, image, audio, video, or sensor data.","importance":"Many real decisions do not rest on text alone.","boundary":"Each modality has its own measurement errors, timelines, and provenance; joint encoding does not remove these boundaries.","keywords":["image","audio","video","sensor","modalities"],"related":["vision-language-models","production-genai","embeddings"]},{"id":"vision-language-models","number":54,"title":"Vision-Language Models","category":"Modality & Production","summary":"Models for joint image and language tasks.","definition":"Vision-language models encode visual inputs and language in a shared or coupled model space, for example for captioning, visual questions, or document analysis.","importance":"They open up screenshots, diagrams, scans, and real scenes for language workflows.","boundary":"An image description is a perception proposal; it is not automatically a provable source record.","keywords":["vlm","vision encoder","image understanding","ocr"],"related":["multimodal-ai","structured-outputs","hallucination-detection","visual-encoder","encoder-free-multimodal"]},{"id":"ai-coding-agents","number":55,"title":"AI Coding Agents","category":"Modality & Production","summary":"Agents that read, change, and test code and operate development tools.","definition":"Coding agents connect language models with repository search, editors, shell, build, tests, and version control.","importance":"They can work on multi-step software tasks over real artifacts.","boundary":"Code generation is not code correctness; rights, scope, tests, review, and reversible changes remain necessary.","keywords":["coding assistant","repository","tests","developer tools"],"related":["ai-agents","tool-calling","mcp","production-genai"]},{"id":"production-genai","number":56,"title":"Production GenAI Systems","category":"Modality & Production","summary":"The whole operable service around models, data, tools, and governance.","definition":"A production GenAI system includes not only a model but data and model versions, APIs, identities, queues, monitoring, tests, rollback, cost, safety boundaries, and human processes.","importance":"Most real errors arise at interfaces and operating states, not only in the forward pass.","boundary":"A demo with good answers is not yet a production-ready system.","keywords":["production","deployment","reliability","operations","governance"],"related":["observability","llm-serving","ai-safety","agentic-workflows"]},{"id":"json","number":57,"title":"JSON","category":"Fundamentals","summary":"A text format for structured values such as objects and arrays.","definition":"JSON represents objects, arrays, strings, numbers, booleans, and null with a small, language-independent syntax.","importance":"Model examples, configuration, and tool responses often use JSON because software can parse its structure.","boundary":"Valid JSON syntax does not prove that a record has the fields or meaning required by a training task.","keywords":["object","array","structured data","serialization"],"matchTerms":["json"],"related":["jsonl","structured-outputs","dataset"],"reference":"https://www.rfc-editor.org/rfc/rfc8259"},{"id":"jsonl","number":58,"title":"JSONL","category":"Fundamentals","summary":"One complete JSON value per line, commonly one example per line.","definition":"A JSON Lines file is UTF-8 text in which each non-empty line is independently valid JSON. A .jsonl extension is customary.","importance":"Examples can be read and validated one record at a time without loading one large JSON array.","boundary":"A syntactically valid line can still be the wrong kind of training example; schema and content checks remain necessary.","keywords":["json lines","line-delimited json","training examples"],"matchTerms":["json lines","jsonl"],"related":["json","dataset","sft","dpo"],"reference":"https://jsonlines.org/"},{"id":"checkpoint","number":59,"title":"Checkpoint","category":"Training & Adaptation","summary":"A saved model state from a particular point in a run.","plain":"A saved copy of the model at one moment of training, like a save point in a game.","definition":"A checkpoint stores model parameters and identifying metadata for a training step. In this app a resumable checkpoint also holds the optimizer moments, the step, the data position and the random state (the learning rate is recomputed from the step); weights-only 'best' checkpoints can be used but not resumed.","importance":"It lets you inspect or compare earlier model states instead of treating the last step as the only result.","boundary":"A checkpoint is not by itself proof of model quality and is not necessarily a portable inference package.","keywords":["weights","training step","resume training","model state"],"matchTerms":["checkpoints"],"related":["training","dataset","validation-loss","llm-inference"],"reference":"https://docs.pytorch.org/tutorials/beginner/saving_loading_models"},{"id":"dataset","number":60,"title":"Dataset","category":"Training & Adaptation","summary":"A collection of examples prepared for training or evaluation.","definition":"A dataset selects and organizes examples, their expected targets, and often separate training and validation portions.","importance":"Its source, coverage, and split determine what a model can learn and what a reported loss actually measures.","boundary":"A large or well-formatted dataset is not automatically representative, clean, or free of overlap between splits.","keywords":["training data","validation split","examples","corpus"],"matchTerms":["datasets"],"related":["jsonl","tokenization","training","validation-loss"]},{"id":"gpt-2","number":61,"title":"GPT-2","category":"Fundamentals","summary":"An OpenAI language model and a widely reused byte-level BPE tokenizer.","definition":"GPT-2 is a Transformer language model introduced in 2019. Its released tokenizer maps UTF-8 bytes to symbols and applies learned byte-pair merges.","importance":"Its published vocabulary gives this app a fixed comparison point beside tokenizers trained on your own text.","boundary":"Selecting the GPT-2 tokenizer does not turn a different model into GPT-2 or transfer GPT-2's learned weights.","keywords":["gpt2","openai","published vocabulary","byte-level bpe"],"matchTerms":["GPT2"],"related":["byte-level","tokenization","transformers"],"reference":"https://openai.com/index/better-language-models/"},{"id":"training","number":62,"title":"Training","category":"Training & Adaptation","summary":"Repeatedly adjusting model parameters from examples and a loss signal.","definition":"In each training step the model predicts, a loss measures error, gradients are computed, and an optimizer updates parameters.","importance":"Watching steps and loss connects model behavior to the data and objective that shaped it.","boundary":"A falling training loss does not guarantee improvement on unseen examples or useful answers.","keywords":["gradient","optimizer","training run","parameter update"],"matchTerms":["training runs"],"related":["dataset","loss","validation-loss","optimizer","checkpoint"],"reference":"https://docs.pytorch.org/tutorials/beginner/basics/optimization_tutorial"},{"id":"byte-level","number":63,"title":"Byte-Level","category":"Fundamentals","summary":"A tokenizer representation built from bytes rather than a fixed list of characters.","definition":"Byte-level tokenization begins with the bytes of encoded text; byte-level BPE can then merge frequent byte sequences into larger tokens.","importance":"A byte base can represent text outside a small character vocabulary without an unknown-character token.","boundary":"Byte-level BPE does not mean every final token is exactly one byte or one visible character.","keywords":["utf-8","bytes","byte pair encoding","bpe"],"matchTerms":["byte level","byte-level bpe"],"related":["gpt-2","tokenization"],"reference":"https://cdn.openai.com/better-language-models/language-models.pdf"},{"id":"loss","number":64,"title":"Loss","category":"Training & Adaptation","summary":"A number that measures error under a chosen training objective.","definition":"A loss function compares predictions with targets. During training, its gradients guide parameter updates; next-token models commonly use cross-entropy loss.","importance":"The loss curve shows whether the model is fitting the task it was given.","boundary":"Lower training loss is not a direct measure of truth, usefulness, or performance on new data.","keywords":["cross entropy","objective","error","training curve"],"matchTerms":["losses","training loss"],"related":["training","validation-loss","optimizer"],"reference":"https://docs.pytorch.org/tutorials/beginner/basics/optimization_tutorial"},{"id":"validation-loss","number":65,"title":"Validation Loss","category":"Training & Adaptation","summary":"Loss measured on held-out examples without updating model weights.","definition":"Validation loss applies the training objective to a separate validation split while leaving parameters unchanged. It helps compare checkpoints and detect a gap from training loss.","importance":"It gives a limited check of generalization beyond the examples used for updates.","boundary":"Repeated choices based on the same validation split can overfit that split; a low value is not a guarantee of good answers. Values are only comparable between runs that share a tokenizer and a validation split.","keywords":["val loss","held-out data","generalization","overfitting"],"matchTerms":["validation losses","validation-loss","val loss","val-loss"],"related":["loss","training","dataset","checkpoint"],"reference":"https://docs.pytorch.org/tutorials/beginner/introyt/trainingyt"},{"id":"forward-pass","number":66,"title":"Forward Pass","category":"Fundamentals","summary":"One run of the model from the input tokens to scores for the next token.","plain":"One complete run through the model: text goes in, a guess for the next word comes out.","definition":"In a forward pass, the token IDs of the context go through the embedding and then through every transformer block in order. For generation, the output head reads the last position and scores every vocabulary token; each new token needs another pass. With a KV cache, later passes run only the newest token and read the earlier keys and values from the cache.","importance":"It is the unit of work behind every predicted token. The Brain Explorer draws one pass as token lanes running through the blocks.","boundary":"A pass computes scores but does not choose a token; sampling does that afterwards. No weights change during inference.","keywords":["inference","forward","next token","token lanes"],"related":["llm-inference","transformer-block","output-head","sampling"]},{"id":"transformer-block","number":67,"title":"Transformer Block","category":"Fundamentals","summary":"One repeated layer of a transformer: attention, then an MLP, each added to the residual stream.","plain":"One floor of the model. The text passes through many such floors, one after another.","definition":"A block normalizes its input, lets its attention heads read from other positions and adds their result to the residual stream. It then normalizes again, applies the MLP at every position and adds that result too. A model stacks many blocks with the same structure but their own weights and runs them one after another.","importance":"Blocks give the model its depth: each one can refine the state that every later block reads.","boundary":"A block is not a head: the heads of multi-head attention work side by side inside one block, while blocks follow one another. The number of blocks alone says nothing about quality.","keywords":["layer","decoder block","n_layer","pre-norm","depth"],"matchTerms":["transformer layer","decoder block"],"related":["transformers","attention-head","multi-head-attention","mlp","residual-stream","layer-norm"],"reference":"https://arxiv.org/abs/1706.03762"},{"id":"attention-head","number":68,"title":"Attention Head","category":"Fundamentals","summary":"One of several attention computations that run side by side inside a transformer block.","plain":"One of several readers in a block; each reader can look for a different kind of connection.","definition":"Each head has its own query projection onto a slice of the vector width (head_dim dimensions). In classic multi-head attention (GPT-2, your own checkpoints) every head also has its own key and value projections. With grouped-query attention (Llama, Qwen2) there are fewer key/value heads than query heads: several query heads share one key and value. Each query head still computes its own attention weights, so heads can weigh the same positions differently; their outputs are joined and projected back into the residual stream.","importance":"Heads let one block follow several relationships at once. The Brain Explorer shows the weights of one recorded head; for an external recording, the Head Finder compares all heads of a recorded round.","boundary":"Heads are not blocks: all heads of a block work in parallel, blocks work in sequence. One head's weights show only part of what the whole attention layer adds.","keywords":["head","n_head","multi-head","query key value","head_dim"],"related":["multi-head-attention","attention","transformer-block","causal-mask"],"reference":"https://arxiv.org/abs/1706.03762"},{"id":"causal-mask","number":69,"title":"Causal Mask","category":"Fundamentals","summary":"The rule that a position may attend only to itself and to earlier positions.","plain":"The rule that a word may only look back, never ahead at words that come later.","definition":"Before the softmax, the attention scores for later positions are set to minus infinity, so their weights become exactly zero. In an attention matrix, this leaves the area above the diagonal empty.","importance":"It lets a model learn next-token prediction at all positions at once without seeing the tokens it has to predict.","boundary":"A masked cell is not a learned weight of zero; that position cannot be reached at all. The mask says nothing about which earlier positions matter.","keywords":["mask","autoregressive","triangular","future tokens","attention matrix"],"matchTerms":["causal masking","causal attention"],"related":["attention","attention-head","softmax","positional-encoding"],"reference":"https://arxiv.org/abs/1706.03762"},{"id":"mlp","number":70,"title":"MLP (Feed-Forward)","category":"Fundamentals","summary":"The part of a transformer block that processes each position on its own.","plain":"The part of a block where each word is worked on by itself, without looking at the others.","definition":"The MLP widens a position's vector, applies a non-linear function and projects it back to the original width; the result is added to the residual stream. In the GPT-2-style models trained in this app, it widens by a factor of four and uses GELU. Llama and Qwen2 use a gated MLP (SwiGLU): two parallel up-projections, one passed through SiLU, are multiplied elementwise and then projected down; the widened size is intermediate_size.","importance":"Attention moves information between positions; the MLP transforms it within each position and holds a large share of the parameters.","boundary":"The MLP does not read from other positions. Within one block, every position uses the same MLP weights.","keywords":["feed forward","multilayer perceptron","gelu","hidden layer","position-wise"],"matchTerms":["MLP","feed-forward","feed-forward network","feed-forward layer","multilayer perceptron"],"related":["transformer-block","residual-stream","attention","transformers"],"reference":"https://arxiv.org/abs/1706.03762"},{"id":"residual-stream","number":71,"title":"Residual Stream","category":"Fundamentals","summary":"The running vector of each position that every block reads from and adds to.","plain":"A running note for each word that every block reads and adds its own result to.","definition":"In GPT-2-style models, each position starts with its token embedding plus its position embedding. Attention and the MLP compute contributions that are added to this vector instead of replacing it (residual connections), so the stream carries information through all blocks to the final normalization (LayerNorm, or RMSNorm in Llama/Qwen2).","importance":"The additions make deep models trainable. They are why the Brain Explorer draws a grey line with plus nodes through all blocks.","boundary":"The stream is a vector with many dimensions; the few components a visualization shows are only a slice, and a single number rarely has a fixed meaning.","keywords":["residual connection","skip connection","hidden state","addition"],"matchTerms":["residual connection","skip connection"],"related":["transformer-block","mlp","attention","layer-norm","embeddings"],"reference":"https://arxiv.org/abs/1512.03385"},{"id":"layer-norm","number":72,"title":"LayerNorm","category":"Fundamentals","summary":"Rescales one position's vector to a standard mean and spread before the next computation.","plain":"Brings the numbers for a word back to a common size, so the next step can work with them.","definition":"Layer normalization subtracts the mean of a vector's components, divides by their standard deviation and then applies a learned scale and shift. Pre-norm blocks apply it before attention and before the MLP; a final LayerNorm prepares the last state for the output head. Llama and Qwen2 use RMSNorm in the same places, which only divides by the root mean square and applies a learned scale, with no mean subtraction and no shift.","importance":"Stable value ranges make training reliable, even as the residual stream grows from block to block.","boundary":"LayerNorm works within the vector of a single position and does not mix positions. Because of the learned scale and shift, its output is not always exactly mean 0 and variance 1.","keywords":["normalization","ln_f","pre-norm","mean","variance"],"matchTerms":["layer norm","layer normalization","layer normalisation"],"related":["transformer-block","residual-stream","output-head"],"reference":"https://arxiv.org/abs/1607.06450"},{"id":"output-head","number":73,"title":"Output Head","category":"Fundamentals","summary":"The final projection that turns the last state into one score per vocabulary token.","plain":"The last step: it gives every possible next word a score.","definition":"After the final normalization (LayerNorm, or RMSNorm in Llama/Qwen2), the output head multiplies a position's state by a matrix with one row per vocabulary token; during generation, only the last position is read. The result is the logits. Many GPT-style models tie this matrix to the token embedding; this app offers that as a model option.","importance":"This is where the model's internal vector becomes a prediction about concrete tokens.","boundary":"Despite its name, the output head is not an attention head. It scores tokens but does not choose one; sampling does that.","keywords":["lm head","unembedding","vocabulary","weight tying","projection"],"matchTerms":["LM head","language model head","unembedding"],"related":["logits","softmax","layer-norm","embeddings","attention-head"]},{"id":"logits","number":74,"title":"Logits","category":"Fundamentals","summary":"Raw, unnormalized scores for every possible next token.","plain":"The raw scores for every possible next word, before they are turned into chances.","definition":"The output head produces one real number per vocabulary token. Higher means more likely, but the values can be negative and do not add up to a fixed total; softmax turns them into probabilities.","importance":"Temperature and filters act on the logits, so they are the starting point of every sampling decision.","boundary":"A logit is not a probability, and differences between logits matter more than their absolute values.","keywords":["score","unnormalized","vocabulary","output head"],"related":["output-head","softmax","temperature","sampling"]},{"id":"softmax","number":75,"title":"Softmax","category":"Fundamentals","summary":"Turns a list of scores into probabilities that are positive and sum to one.","plain":"Turns the scores into chances between 0 and 100% that add up to exactly 100%.","definition":"Softmax applies the exponential function to each score and divides by the sum of all exponentials. Larger scores get larger shares, and adding the same constant to every score changes nothing. Transformers use it for attention weights and for next-token probabilities.","importance":"It makes scores comparable as shares, which is what probability bars and attention arcs show.","boundary":"Softmax probabilities describe the model's distribution, not the chance that a statement is true.","keywords":["probability","normalization","exponential","distribution"],"related":["logits","attention","temperature","causal-mask"],"reference":"https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.softmax.html"},{"id":"sampling","number":76,"title":"Sampling","category":"Inference & Operations","summary":"Choosing the next token from the model's probability distribution.","plain":"Picking the next word by chance, where likely words come up more often.","definition":"After temperature and optional filters such as top-k and top-p, the next token is drawn at random in proportion to the remaining probabilities. Greedy decoding instead always takes the most likely token, which corresponds to temperature 0.","importance":"The same model and prompt can produce different continuations; the sampling settings control how varied and how risky they are.","boundary":"Sampling changes neither what the model has learned nor its scores. A drawn token is not automatically the best or a correct one.","keywords":["decoding","greedy","random","next token","distribution"],"matchTerms":["greedy decoding"],"related":["temperature","top-k","top-p","seed","softmax","logits"]},{"id":"temperature","number":77,"title":"Temperature","category":"Inference & Operations","summary":"Divides the logits before sampling: below 1 the distribution gets peakier, above 1 flatter.","plain":"How adventurous the choice is: low means safe and predictable, high means more surprising.","definition":"The logits are divided by the temperature before the softmax. Below 1 the distribution gets peakier and likely tokens become even more likely; above 1 it gets flatter and unlikely tokens gain probability; temperature 0 always takes the most likely token.","importance":"It is the main control between predictable and varied text.","boundary":"Temperature changes the selection distribution, not the model's scores or its knowledge.","keywords":["sampling temperature","creativity","greedy","distribution"],"related":["sampling","logits","softmax","top-k","top-p"]},{"id":"top-k","number":78,"title":"Top-k","category":"Inference & Operations","summary":"A sampling filter that keeps only the k most likely tokens.","plain":"Only the k most likely next words stay in the draw; all others are left out.","definition":"All tokens outside the k highest scores are removed before drawing, and the remaining probabilities are normalized again. In this app, k = 0 switches the filter off; it is applied before top-p.","importance":"It cuts off the long tail of unlikely tokens that can derail a continuation.","boundary":"A fixed k ignores how the probability is spread: sometimes k tokens are too many, sometimes too few.","keywords":["filter","candidates","truncation","top k"],"matchTerms":["top-k sampling"],"related":["top-p","sampling","temperature"]},{"id":"top-p","number":79,"title":"Top-p","category":"Inference & Operations","summary":"A sampling filter that keeps the smallest set of tokens that together reach probability p.","plain":"Only the most likely words stay in the draw, until together they reach the chance p.","definition":"Tokens are sorted by probability and kept until their cumulative probability reaches p; the rest are removed before drawing. The most likely token always stays, and p = 1 keeps every token.","importance":"Unlike top-k, it adapts to the distribution: few candidates when the model is confident, more when it is unsure.","boundary":"Top-p limits the probability mass that is drawn from, not the quality of the continuation.","keywords":["nucleus","cumulative probability","filter","top p"],"matchTerms":["nucleus sampling","top-p sampling"],"related":["top-k","sampling","temperature","softmax"],"reference":"https://arxiv.org/abs/1904.09751"},{"id":"seed","number":80,"title":"Seed","category":"Inference & Operations","summary":"The start value of a pseudo-random number generator.","plain":"The starting number for the random choices. The same seed gives the same choices again.","definition":"The same seed with the same code, model, settings and hardware repeats the same random choices, for example the tokens drawn during sampling or the order of examples in training.","importance":"It makes experiments repeatable and comparable.","boundary":"A seed does not make a result correct, and equal seeds with different inputs or settings lead to different draws. Some GPU operations are not fully deterministic even with a fixed seed.","keywords":["random","reproducibility","deterministic","generator"],"matchTerms":["random seed"],"related":["sampling","temperature","training"],"reference":"https://docs.pytorch.org/docs/stable/notes/randomness.html"},{"id":"context","number":81,"title":"Context","category":"Fundamentals","summary":"The tokens a model actually processes in one pass: the prompt plus everything generated so far.","plain":"Everything the model can see right now: your input plus the words it has already written.","definition":"During generation the context grows by one token per round: first the input tokens, then the tokens chosen in earlier rounds. It can grow up to the context window; beyond that the app's own models keep only the most recent block_size tokens, and models with a KV cache compute only the new token and reuse the cached keys and values for the rest.","importance":"The next token depends only on the context: if one token in it changes, every later lane can change.","boundary":"The context is not the context window. A longer context is not automatically used better, and text beyond the context window never reaches the model.","keywords":["prompt","input","sequence","tokens so far"],"related":["context-windows","tokenization","forward-pass","kv-cache"]},{"id":"token-embedding","number":82,"title":"Token Embedding","category":"Fundamentals","summary":"The learned vector a language model looks up for each token ID before the first block.","plain":"Each token is swapped for a long list of numbers that the model has learned for it.","definition":"The embedding table holds one row of n_embd numbers for every vocabulary token. Each token ID in the context selects its row; in GPT-2-style models the vector of its position is added, and the sum enters the first block.","importance":"It turns discrete IDs into numbers the blocks can compute with; tokens used in similar ways can end up with similar vectors.","boundary":"The vector is learned and has no fixed dictionary meaning; on its own it knows nothing about the token's position or neighbours.","keywords":["wte","embedding table","lookup","token vector"],"matchTerms":["embedding table"],"related":["embeddings","tokenization","positional-encoding","residual-stream"]},{"id":"diffusion-model","reference":"https://arxiv.org/abs/2006.11239","number":83,"title":"Diffusion Model","category":"Modality & Production","summary":"A generator that learns to reverse noise corruption through repeated updates.","plain":"An image generator that starts from pure noise and cleans it up step by step.","definition":"The Vision learning model predicts added noise from a noisy image, a diffusion timestep and a caption. Generation starts from noise and uses the learned predictions in a sequence of scheduler updates.","importance":"Explains why image generation has intermediate states across the whole image.","boundary":"A denoising update changes the sample, while a training update changes the weights.","keywords":["diffusion model","diffusion models","DDPM"],"matchTerms":["diffusion model","diffusion models","DDPM"],"related":["noise-prediction","noise-schedule","latent-diffusion"]},{"id":"diffusion-timestep","reference":"https://arxiv.org/abs/2006.11239","number":84,"title":"Diffusion Timestep","category":"Modality & Production","summary":"An index identifying the noise level supplied to the denoising model.","plain":"A number that tells the model how noisy the picture still is at this step.","definition":"The small Vision model receives a timestep alongside the noisy image. During training each example can use a different timestep; generation walks through its diffusion schedule.","importance":"Separates noise level from the number of optimizer updates.","boundary":"A timestep is not elapsed wall-clock time or the training-step counter.","keywords":["diffusion timestep","diffusion timesteps"],"matchTerms":["diffusion timestep","diffusion timesteps"],"related":["noise-schedule","diffusion-model"]},{"id":"noise-schedule","reference":"https://arxiv.org/abs/2006.11239","number":85,"title":"Noise Schedule","category":"Modality & Production","summary":"The sequence of noise strengths used by a diffusion process.","plain":"The plan for how much noise is in the picture at each step.","definition":"The Vision learning model stores noise coefficients for its diffusion timesteps. Training combines an image with noise using those coefficients; generation uses the corresponding reverse update.","importance":"Connects a timestep to the amount of corruption in the input.","boundary":"A noise schedule differs from a learning-rate schedule, which controls weight updates.","keywords":["noise schedule","noise schedules"],"matchTerms":["noise schedule","noise schedules"],"related":["diffusion-timestep","noise-prediction"]},{"id":"image-caption","number":86,"title":"Image Caption","category":"Modality & Production","summary":"A text description paired with an image for conditioning a model.","definition":"Vision pairs every training image with a caption of its visible content. The small model encodes UTF-8 bytes and averages learned byte embeddings to condition noise prediction.","importance":"Consistent descriptions help connect visual variation to text.","boundary":"This small byte encoder differs from the pretrained starter's text encoder.","keywords":["image caption","image captions","caption","captions"],"matchTerms":["image caption","image captions","caption","captions"],"related":["dataset","tokenization","diffusion-model"]},{"id":"noise-prediction","reference":"https://arxiv.org/abs/2006.11239","number":87,"title":"Noise Prediction","category":"Modality & Production","summary":"Estimating the noise component in a corrupted image or latent state.","plain":"The model's guess of which part of the picture is only noise.","definition":"In the small Vision trainer, random noise is added to an image. The model predicts that noise from the corrupted image, caption and timestep; MSE compares the prediction to the known noise.","importance":"The training target is available automatically because the code creates it.","boundary":"The prediction is noise, not a finished image or a semantic object label.","keywords":["noise prediction","noise predictions","predicted noise","target noise"],"matchTerms":["noise prediction","noise predictions","predicted noise","target noise"],"related":["diffusion-model","mse","diffusion-timestep"]},{"id":"mse","reference":"https://huggingface.co/docs/diffusers/tutorials/basic_training","number":88,"title":"Mean Squared Error","category":"Modality & Production","summary":"The average squared difference between predicted and target values.","definition":"MSE squares each prediction error and averages the results. Vision applies it to predicted and added noise; batch training loss and held-out validation loss use this objective.","importance":"Provides a scalar objective from many numerical prediction errors.","boundary":"Its scale depends on the target representation; it is not a perceptual image-quality score.","keywords":["MSE","mean squared error"],"matchTerms":["MSE","mean squared error"],"related":["loss","noise-prediction","validation-loss"]},{"id":"latent-diffusion","reference":"https://arxiv.org/abs/2112.10752","number":89,"title":"Latent Diffusion","category":"Modality & Production","summary":"Diffusion performed in a compressed image representation rather than directly in RGB pixels.","plain":"The noise is cleaned up in a small, compressed version of the image, which is faster.","definition":"The pretrained Vision starter updates a latent state. Its VAE decoder turns that state into image previews. The app's own small trainer instead updates pixel-space states.","importance":"Explains why a visible preview differs from the state the denoiser processes.","boundary":"Latent components are not RGB pixels or individually named objects.","keywords":["latent diffusion","latent state","latent states"],"matchTerms":["latent diffusion","latent state","latent states"],"related":["diffusion-model","vae","noise-prediction"]},{"id":"vae","reference":"https://arxiv.org/abs/1312.6114","number":90,"title":"Variational Autoencoder","category":"Modality & Production","summary":"An encoder-decoder model that learns a probabilistic compressed representation.","plain":"Packs an image into a small, compressed form and unpacks it back into a picture.","definition":"A VAE encodes images into a latent distribution and decodes latent representations back into images. Vision uses the starter model's decoder to display recorded latent states.","importance":"Connects compressed diffusion states to the pictures users can inspect.","boundary":"Decoding can lose or change details; the tiny pixel-space trainer has no VAE.","keywords":["VAE","variational autoencoder","variational autoencoders"],"matchTerms":["VAE","variational autoencoder","variational autoencoders"],"related":["latent-diffusion","embeddings"]},{"id":"u-net","reference":"https://arxiv.org/abs/1505.04597","number":91,"title":"U-Net","category":"Modality & Production","summary":"A downsampling and upsampling network with skip connections between matching resolutions.","plain":"The network that looks at the noisy image both in fine detail and as a rough overview.","definition":"U-Net-style denoisers combine broader features from reduced-resolution layers with finer features passed through skip connections. Vision's small denoiser uses a compact down-and-up convolutional path with a skip connection.","importance":"Explains how image features at different spatial scales can be combined.","boundary":"A U-Net need not include attention, and its stages are not Transformer blocks.","keywords":["U-Net","U-Nets","UNet"],"matchTerms":["U-Net","U-Nets","UNet"],"related":["diffusion-model","noise-prediction"]},{"id":"classifier-free-guidance","reference":"https://arxiv.org/abs/2207.12598","number":92,"title":"Classifier-Free Guidance","category":"Modality & Production","summary":"Combining conditioned and unconditioned predictions to adjust prompt influence.","plain":"Decides how closely the picture should follow your text description.","definition":"The pretrained Vision path uses a guidance scale to combine denoiser predictions with and without text conditioning. The small pixel-space model has no classifier-free guidance control.","importance":"Connects the guidance setting to text conditioning during generation.","boundary":"Higher guidance does not necessarily improve quality; it is not sampling temperature.","keywords":["classifier-free guidance","guidance scale"],"matchTerms":["classifier-free guidance","guidance scale"],"related":["latent-diffusion","image-caption","noise-prediction"]},{"id":"cross-attention","reference":"https://arxiv.org/abs/2112.10752","number":93,"title":"Cross-Attention","category":"Modality & Production","summary":"Attention in which image positions read the text encoder's token vectors.","plain":"Here the parts of the picture look at the words of your description.","definition":"In the Stable Diffusion U-Net, cross-attention layers let every spatial position of the latent image weigh the prompt tokens: queries come from image features, keys and values from the text encoder's output. The Vision Explorer shows the weights of one recorded cross-attention layer per denoising step.","importance":"Shows where in the image the prompt can steer the prediction during denoising.","boundary":"The weights show what one layer reads from the prompt, not why something is drawn; other layers and heads also contribute.","keywords":["cross-attention","cross attention"],"matchTerms":["cross-attention","cross attention"],"related":["attention","u-net","text-encoder","classifier-free-guidance"]},{"id":"scheduler","reference":"https://arxiv.org/abs/2202.09778","number":94,"title":"Scheduler (Sampler)","category":"Modality & Production","summary":"The rule that turns each step's predicted noise into the next, less noisy state.","plain":"The rule that removes a bit of the guessed noise at each step.","definition":"A diffusion scheduler sets the sequence of timesteps and, at every step, combines the current noisy state with the model's noise prediction into the next state. The pretrained Vision models use PNDM; each recorded step shows the state before, the predicted noise and the change.","importance":"Explains why the same model can give different images with another scheduler or step count.","boundary":"A scheduler has no trained weights of its own and does not decide what is drawn.","keywords":["scheduler","sampler","PNDM"],"matchTerms":["scheduler","sampler","PNDM"],"related":["noise-schedule","diffusion-timestep","noise-prediction","diffusion-model"]},{"id":"text-encoder","reference":"https://arxiv.org/abs/2103.00020","number":95,"title":"Text Encoder (CLIP)","category":"Modality & Production","summary":"The network that turns the prompt's tokens into vectors the image denoiser reads.","plain":"Translates your text description into numbers that the image generator can read.","definition":"Stable Diffusion splits the prompt into at most 77 tokens, pads the rest and runs them through a CLIP text encoder. Its output vectors are the keys and values of the U-Net's cross-attention at every denoising step.","importance":"Connects the typed prompt to what the denoiser can read while it removes noise.","boundary":"The text encoder stays unchanged during image generation and does not produce text.","keywords":["text encoder","CLIP"],"matchTerms":["text encoder","CLIP"],"related":["tokenization","cross-attention","latent-diffusion","classifier-free-guidance"]},{"id":"visual-encoder","reference":"https://arxiv.org/abs/2502.14786","number":96,"title":"Visual Encoder","category":"Modality & Production","summary":"A pretrained image network that turns a picture into tokens a language model can read.","definition":"In encoder-based vision-language models, a vision transformer such as CLIP or SigLIP cuts the image into patches, lets them attend to each other in both directions and passes the result through an adapter and a projector into the language model as visual tokens. It brings a visual prior learned from large collections of image-text pairs.","importance":"The language model can use images from the first training step instead of learning to see from raw pixels.","boundary":"This is the image side of models that understand images. In Stable Diffusion, the CLIP text encoder reads the prompt, not an image.","keywords":["visual encoder","image encoder","ViT","SigLIP","encoder-based"],"matchTerms":["visual encoder","image encoder"],"related":["vision-language-models","image-patch-token","encoder-free-multimodal","text-encoder","attention"]},{"id":"encoder-free-multimodal","reference":"https://arxiv.org/abs/2609.35457","number":97,"title":"Encoder-Free Multimodal Model","category":"Modality & Production","summary":"A language model that reads image patches directly, without a separate pretrained image network.","definition":"Encoder-free models project raw image patches straight into the language model, which has to learn visual features itself. In a controlled comparison (Chen et al., 2026) they reached a higher multimodal loss than encoder-based models at the measured sizes, but their loss fell faster with more compute; their early layers took over the work a visual encoder would do.","importance":"Shows a different design: one network for image and text instead of an image encoder plus a language model.","boundary":"The predicted catch-up at about 10^22 training FLOPs is an extrapolation of fitted scaling laws, far beyond anything this app trains.","keywords":["encoder-free","encoderless","native multimodal","Fuyu","EVE"],"matchTerms":["encoder-free"],"related":["visual-encoder","image-patch-token","scaling-laws","vision-language-models","transformer-block"]},{"id":"image-patch-token","reference":"https://arxiv.org/abs/2010.11929","number":98,"title":"Image Patch Token","category":"Modality & Production","summary":"A small square of an image turned into one vector, so a transformer can treat the image as a sequence.","definition":"The image is cut into fixed-size patches, for example 16×16 pixels; each patch is flattened and projected to the model width, giving one token per patch. Within one image these tokens are often allowed to attend to each other in both directions, unlike text tokens under a causal mask.","importance":"Lets images use the same token and attention machinery as text.","boundary":"A patch token holds pixels, not a word; it only means what the network learns from it.","keywords":["patch","image patch","patch embedding","visual token"],"matchTerms":["image patch","patch token","visual token"],"related":["tokenization","token-embedding","causal-mask","visual-encoder","encoder-free-multimodal"]},{"id":"scaling-laws","reference":"https://arxiv.org/abs/2203.15556","number":99,"title":"Scaling Laws","category":"Training & Adaptation","summary":"Fitted rules for how loss falls as model size, training data and compute grow.","definition":"Scaling laws fit validation loss as a power law of training compute, model size and number of tokens, using many smaller runs. The Chinchilla method trains models of different sizes at the same compute budget (IsoFLOP profiles) to find the compute-optimal split between parameters and tokens.","importance":"Lets researchers predict large runs from small ones and compare designs at equal compute.","boundary":"A fitted law describes the measured range; anything beyond it is an extrapolation and can be wrong.","keywords":["scaling law","Chinchilla","compute-optimal","IsoFLOP"],"matchTerms":["scaling law","scaling laws","compute-optimal"],"related":["training","loss","validation-loss","encoder-free-multimodal"]},{"id":"rgb444","number":100,"title":"RGB444 colour tokens","category":"Modality & Production","summary":"A palette of 4,096 colours: one token ID per pixel colour.","plain":"Each colour gets a number. The number says which colour, not where it is.","definition":"RGB444 uses four bits for each red, green and blue channel. With channel values 0–15, ID = red × 256 + green × 16 + blue. Display channels multiply those values by 17. The shapes learner predicts these colour IDs per image row.","importance":"Makes image tokenization directly reversible and inspectable without a learned image codebook.","boundary":"A colour token contains neither location nor object identity. A larger palette does not shorten the image sequence.","keywords":["RGB444","colour token","palette"],"matchTerms":["RGB444","colour token"],"related":["tokenization","image-caption"]},{"id":"teacher-forcing","number":101,"title":"Teacher forcing","category":"Training & Adaptation","summary":"Training predicts the next target using the correct preceding data rather than its own generated output.","plain":"During practice the model sees the correct previous step. Later it has to use its own answer.","definition":"In the shapes row model, each target row is predicted from the caption, position, recurrent state and correct preceding image row. Generation instead feeds back the predicted preceding row. The training loss is cross-entropy over 4,096 colour classes per pixel.","importance":"Explains why a falling training loss can coexist with mistakes accumulating during generation.","boundary":"It does not guarantee correct free-running images. Pixels within the same generated row are conditionally independent.","keywords":["teacher forcing","row generation"],"matchTerms":["teacher forcing"],"related":["rgb444","training","loss"]},{"id":"learning-rate","number":102,"title":"Learning Rate","category":"Training & Adaptation","summary":"How far the optimizer moves the weights in one step.","plain":"How big each correction is. Too big and training jumps around, too small and it barely moves.","definition":"The learning rate scales every change the optimizer makes to the weights. On the Training page it rises during a short warmup and then falls along a cosine curve to a tenth of its peak; on the Shapes page it falls along a cosine curve to a tenth or stays constant.","importance":"Too high a rate makes the loss jump or diverge; too low a rate wastes steps.","boundary":"There is no universally right value. It interacts with batch size, optimizer and model size, and a good value for one run is only a starting point for another.","keywords":["learning rate","warmup","cosine schedule","lr"],"matchTerms":["learning rate","learning rates"],"related":["optimizer","training-step","batch-size","adamw","training"]},{"id":"batch-size","number":103,"title":"Batch Size","category":"Training & Adaptation","summary":"How many examples one step averages before the weights change.","plain":"How many examples the model looks at before it corrects itself once.","definition":"A batch is the group of examples whose loss is averaged into one gradient and one update. On the Training page the batch of a step is built from micro batches: the micro batch is what fits on the card at once, and gradient accumulation adds micro batches up to the planned size of the step. On the Shapes page the batch size is the number of pictures per step.","importance":"Larger batches give a steadier gradient per step but cost memory and time; together with the steps they set how much data a run sees.","boundary":"A larger batch is not automatically better, and a different batch size usually needs a different learning rate. The micro batch changes speed and memory, not the size of a step.","keywords":["batch","micro batch","gradient accumulation"],"matchTerms":["batch","batches","batch size","batch sizes"],"related":["training-step","epoch","learning-rate","training"]},{"id":"training-step","number":104,"title":"Step","category":"Training & Adaptation","summary":"One update of the weights: a batch in, a loss, a gradient, one optimizer change.","plain":"One round of practice: the model looks at a batch, measures its error and corrects its weights once.","definition":"A training step runs a batch through the model, computes the loss, propagates the gradient back and lets the optimizer change the weights once. The step counter of a run counts these updates; checkpoints, evaluations and the learning-rate schedule are placed by step.","importance":"Steps are the clock of a training run: run length, save interval and loss curves are measured in them.","boundary":"A step is not an epoch and not a generation step. How much one step learns depends on batch size and learning rate. The denoising steps of image generation are a different counter.","keywords":["training step","optimizer step","update"],"matchTerms":["steps","training step","training steps"],"notIn":["Toerken.","Workflow.","Course.","Learn.ChapterPlain.","Settings.Demo.","Data.Empty.","Runs.Download.","Telemetry.","Vision.Steps","Vision.StepProgress","Onboarding.Tokenizer.Why","Sample.Rail.MinPHint"],"related":["batch-size","epoch","learning-rate","checkpoint","training"]},{"id":"epoch","number":105,"title":"Epoch","category":"Training & Adaptation","summary":"One full pass of the training data through the model.","plain":"Once through all training examples. Ten epochs means each example came up about ten times.","definition":"An epoch is complete when every training example has been used once. With a fixed batch size, epochs ≈ steps × batch size ÷ number of training examples. The Shapes page shows this estimate next to its steps; the Training page counts steps and tokens instead.","importance":"Epochs tell how often the model sees the same data again, which is what drives memorisation on small datasets.","boundary":"The app trains by steps, not by epochs, so an epoch count is an estimate. More epochs on a small set can lower the training loss while validation gets worse.","keywords":["epoch","pass","memorisation"],"matchTerms":["epoch","epochs"],"related":["training-step","batch-size","dataset","validation-loss"]},{"id":"gate","number":106,"title":"Gate","category":"Training & Adaptation","summary":"A learned number between 0 and 1 that mixes two sources into one result.","plain":"Like a switch the model sets for each pixel: how much of the colour comes from the description, how much from itself.","definition":"In the Shapes model every pixel computes a gate g between 0 and 1. The probability of a colour is g × the colour the description names plus (1 − g) × the model's own colour prediction. The Explorer shows g as the share taken from the description; if the description names no colour, the gate is closed (0).","importance":"It shows whether a pixel's colour was taken from the text or came from what the model learned, for example the colour of the row above.","boundary":"A gate is a computed weight, not a decision or an intention of the model. A gate near 100 % says where the colour came from, not that the pixel lies in the right place.","keywords":["gate","gating","pointer-generator","mixture"],"matchTerms":["gate","gates","gating"],"related":["rgb444","teacher-forcing","attention","softmax"]},{"id":"python","number":107,"title":"Python","category":"Inference & Operations","summary":"A programming language used to build many AI tools.","plain":"Python is the language on the computing side of this app; you operate it through the interface.","definition":"Python runs programs. LLM Studio uses Python on its computing side to prepare data and call libraries such as PyTorch. The Windows interface is a separate part of the app.","importance":"This explains why setup prepares a Python environment and packages as well as the application itself.","boundary":"Python by itself is not a trained model. You do not need to write Python to use the learning paths.","keywords":["Python","python"],"matchTerms":["Python"],"related":["pytorch","numpy"],"reference":"https://docs.python.org/3/tutorial/"},{"id":"pytorch","number":108,"title":"PyTorch","category":"Inference & Operations","summary":"A software toolkit for computing with neural networks.","plain":"PyTorch does the number crunching that lets a model learn and calculate outputs.","definition":"PyTorch is a library that can be used from Python. It computes with tensors, provides neural-network building blocks and automatically calculates derivatives used to update weights during training. LLM Studio uses it on its computing side.","importance":"The model is the architecture plus its learned weights; PyTorch is the tool that performs the calculations.","boundary":"PyTorch is not a ready-made chatbot or a specific model. CPU or GPU use depends on installation, hardware and the selected workflow.","keywords":["PyTorch","pytorch"],"matchTerms":["PyTorch"],"related":["python","tensor","cpu","gpu","cuda","optimizer"],"reference":"https://docs.pytorch.org/tutorials/beginner/basics/intro.html"},{"id":"tensor","number":109,"title":"Tensor","category":"Inference & Operations","summary":"An array of numbers with a defined shape.","plain":"A tensor is an organised bundle of numbers: a list, a table or a stack of tables.","definition":"In PyTorch, a tensor may hold a word vector, image colours or layer weights. Its shape gives the size of each axis: 3 × 64 × 64 can mean three colour channels of a 64 × 64 image. Its data type determines how each number is stored.","importance":"Tensors connect the colour grids and vector strips in the explorers to actual numbers inside the model.","boundary":"Shape alone does not tell you what the axes mean. Three axes can represent very different things at different calculation steps.","keywords":["Tensor","tensor","shape","array","dimensions"],"matchTerms":["Tensor","Tensors"],"related":["pytorch","numpy","embeddings","safetensors"],"reference":"https://docs.pytorch.org/tutorials/beginner/basics/tensorqs_tutorial.html"},{"id":"cpu","number":110,"title":"CPU","category":"Inference & Operations","summary":"The computer’s general-purpose main processor.","plain":"The CPU handles everyday computing tasks and can also run model calculations.","definition":"CPU stands for Central Processing Unit. It runs programs, organises data and computes tensor operations. In CPU computation, most model data is held in system memory, called RAM.","importance":"CPU and GPU are two possible places to compute. This helps explain why a task can take longer without a graphics card.","boundary":"PyTorch CPU support does not mean every app feature or an arbitrarily large model is enabled on CPU. Runtime and available RAM limit what is practical.","keywords":["CPU","cpu","RAM","processor"],"matchTerms":["CPU"],"related":["gpu","pytorch"],"reference":"https://docs.pytorch.org/tutorials/beginner/basics/tensorqs_tutorial.html"},{"id":"gpu","number":111,"title":"GPU","category":"Inference & Operations","summary":"A processor designed for many parallel calculations.","plain":"The GPU can do many similar number calculations at once, which helps with training.","definition":"GPU stands for Graphics Processing Unit. Many neural-network operations can be split into numerous parallel calculations. A suitable GPU can accelerate them; model data and intermediate results need room in its video memory, called VRAM.","importance":"This explains why both GPU support and free VRAM matter during training.","boundary":"A GPU does not accelerate every task. Hardware, algorithms and software must be compatible; system RAM and VRAM are different memory pools.","keywords":["GPU","gpu","VRAM","graphics card"],"matchTerms":["GPU"],"related":["cpu","cuda","pytorch"],"reference":"https://docs.pytorch.org/tutorials/beginner/basics/tensorqs_tutorial.html"},{"id":"cuda","number":112,"title":"CUDA","category":"Inference & Operations","summary":"NVIDIA’s software platform for computing on supported NVIDIA GPUs.","plain":"CUDA is one way PyTorch uses a compatible NVIDIA graphics card for calculations.","definition":"CUDA provides programming interfaces and tools for general-purpose computation on NVIDIA GPUs. With a compatible PyTorch setup, tensor operations run on the GPU rather than the CPU. A compatible NVIDIA driver is required.","importance":"The term connects the graphics card, its driver and the installed PyTorch build.","boundary":"CUDA is not a graphics card or a model. A CPU installation does not need it; an arbitrary GPU or a separately installed toolkit version does not guarantee compatibility.","keywords":["CUDA","cuda"],"matchTerms":["CUDA"],"related":["gpu","pytorch","cpu"],"reference":"https://docs.nvidia.com/cuda/cuda-programming-guide/index.html"},{"id":"numpy","number":113,"title":"NumPy","category":"Inference & Operations","summary":"A Python library for numerical arrays and data processing.","plain":"NumPy helps store and process large lists and tables of numbers efficiently.","definition":"NumPy provides multidimensional arrays and numerical operations. This app uses it on the computing side, including data preparation. PyTorch tensors resemble these arrays but can also compute on supported GPUs and track derivatives for training.","importance":"This separates data-processing tools from model libraries: both work with numbers but have different jobs.","boundary":"A NumPy array is not a trained neural network. NumPy and PyTorch are not interchangeable in every operation or memory layout.","keywords":["NumPy","numpy"],"matchTerms":["NumPy"],"related":["python","tensor","pytorch"],"reference":"https://numpy.org/doc/stable/user/whatisnumpy.html"},{"id":"hugging-face-hub","number":114,"title":"Hugging Face Hub","category":"Inference & Operations","summary":"A platform for model files, datasets and their documentation.","plain":"The Hub is a place to download model files that fit your task.","definition":"Hugging Face runs the Hub and develops libraries such as Transformers and Diffusers. A model repository may contain weights, configuration, a tokenizer and a Model Card describing the model. Downloaded models can run locally with suitable software.","importance":"This separates the download source, the software library and the model itself.","boundary":"A Hub listing does not guarantee quality or app compatibility. Check the model description and its terms of use; loading locally is different from calling a hosted API.","keywords":["Hugging Face Hub","hugging-face-hub"],"matchTerms":["Hugging Face Hub"],"related":["transformers-library","diffusers","safetensors"],"reference":"https://huggingface.co/docs/hub/index"},{"id":"transformers-library","number":115,"title":"Transformers (Hugging Face)","category":"Inference & Operations","summary":"A library for loading and running supported model architectures.","plain":"The Transformers library connects suitable model files with the code that computes them.","definition":"Hugging Face Transformers provides model architecture implementations and tools for tokenizers, inputs and outputs. This app uses the library for supported external models. The trained weights are separate files.","importance":"Transformers can refer either to the general architecture or to this particular software library.","boundary":"The library is not itself Qwen or another checkpoint. Library support does not automatically mean every architecture and visualisation is supported in LLM Studio.","keywords":["Transformers (Hugging Face)","transformers-library"],"matchTerms":["Transformers (Hugging Face)","Hugging Face Transformers"],"related":["transformers","pytorch","hugging-face-hub","tokenization"],"reference":"https://huggingface.co/docs/transformers/index"},{"id":"diffusers","number":116,"title":"Diffusers","category":"Inference & Operations","summary":"A library of building blocks for diffusion models.","plain":"Diffusers connects the steps that turn noise into an image.","definition":"Hugging Face Diffusers provides model components, schedulers and pipelines for diffusion methods. The diffusion path on the Vision page uses such components. A pipeline can connect a text encoder, noise predictor, scheduler and image decoder.","importance":"This separates the library, the image model and the individual steps shown in the Vision Explorer.","boundary":"Diffusers is not one particular trained image model. The Shapes learning mode with colour tokens uses a different computation path, not a Diffusers pipeline.","keywords":["Diffusers","diffusers"],"matchTerms":["Diffusers"],"related":["pytorch","hugging-face-hub","diffusion-model"],"reference":"https://huggingface.co/docs/diffusers/index"},{"id":"safetensors","number":117,"title":"Safetensors","category":"Inference & Operations","summary":"A file format for storing tensors, often model weights.","plain":"A Safetensors file holds a model’s numbers, not the whole program around it.","definition":"Safetensors stores named tensors with their shapes, data types and values. Unlike Python Pickle, reading this format does not require executing embedded Python objects. Using the weights still needs a matching architecture, configuration and, depending on the model, a tokenizer.","importance":"This explains why a downloaded weights file is not enough by itself and why checkpoints must match the model type.","boundary":"The format does not establish origin, quality or usage rights. The rest of the model package and its program code also need review.","keywords":["Safetensors","safetensors"],"matchTerms":["Safetensors"],"related":["tensor","pytorch","hugging-face-hub"],"reference":"https://huggingface.co/docs/safetensors/index"}],"revision":"a6faabca957c80d6bc83223d25f6cda0ba14e3901435e609cab7dca71bc273f1","sourceCommit":"3df13173e1cdb99a2e3ce5a2ca9dccd1013091b9","sourceDirty":false,"sourceSha256":"a6faabca957c80d6bc83223d25f6cda0ba14e3901435e609cab7dca71bc273f1","revisionAlgorithm":"sha256-utf8-lf-no-bom"}