Brain Explorer, Wiki & Vision
Alpha 0.2.0 adds a 2D Brain Explorer, a multilingual Wiki and learning tools for image generation and image captioning.
- A reorganized learning pipeline separates source text, tokenization and dataset preparation; the published GPT-2 vocabulary is available as an optional download.
- A multilingual Wiki adds search, related topics and concept explanations in Learn and supported app text, including recognition of word forms and plurals.
- A new 2D Brain Explorer follows context through attention and MLP to probabilities, sampling and appending the next token.
- Recorded generation supports playback, pause and step-by-step inspection. A/B views compare two inputs through the same model.
- Token lanes distinguish input from generated text. Inspect attention, individual heads, vector slices and, for supported external recordings, the KV cache.
- For supported external recordings, compare hypothetical temperature and sampling filters without changing the recorded token choice.
- Inspect supported linked Llama/Qwen2 models in the Brain and Visual Explorer; diagrams reflect RoPE, RMSNorm, GQA and SwiGLU where used by the architecture.
- A curated catalogue supports optional model downloads after licence confirmation. Brain Explorer recordings use binary storage with a SQLite catalogue.
- A dedicated Vision page provides local image generation, model selection, licence-confirmed downloads and CPU/GPU selection for supported models.
- The Vision Explorer compares state before, noise prediction and state after. Feature maps, exact values and heads at one selected cross-attention site connect image grids with text tokens.
- Record new images directly in the Vision Explorer, with progress, cancellation, saved recordings and an explain mode.
- Train a small text-conditioned image model on image-caption pairs. The current pixel learning path uses 32 × 32 images and shows training/validation loss and checkpoint comparisons.
- The image-understanding section compares small captioning models using a pretrained visual encoder, direct patch projection and a causal patch variant through local training runs.
- Additional source-text, SFT and DPO dataset downloads, language filters and consistent licence dialogs with links and licence-type labels.
- Learn and Wiki cover the new image, diffusion, sampling and multimodal concepts in German, English and Romanian.
- Consistent Explorer controls, stable selection lists, improved weight popups, optional GPU status and reduced-motion support.
- More robust Fine-tuning startup, no restoration of scroll positions, and understandable safety-rejection messages instead of technical log output.
- The installer backs up existing program files before updates and restores them on installation failure, preserving personal data.