- scripts/jwash_mcp.py: MCP server (FastMCP/stdio), an HTTP client of the running
J-Wash server so an external LLM can drive an already-loaded model. Tools:
generate, find_token, list_layers, scale_token/replace_token (pure-weights,
layers required), set_intensity, list_edits/reset_edits.
- api/app.py + ui: /api/generate now records the last exchange (surfaced in
/api/status) and broadcasts on /ws when done; the UI shows a "generated via
API/MCP" panel at the top of the chat and an Options "API monitor" toggle that
swaps the 2s status poll for an event-driven refresh.
- sampling: repetition penalty (default 1.0, applied in model_manager._sample),
exposed as a "rep" field in the chat controls; intentionally not exposed
through the MCP.
- ui: remove the redundant "md" chat toggle (already available in Options).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The fit corpus was limited to three hardcoded choices (wikitext, Semantic-Harmless, mixed), and any other id was rejected. Now any HuggingFace dataset id works, and any number of them can be ticked to fit on an equal-parts mix, shuffled.
n_prompts now counts training SEQUENCES (what the fit iterates over) instead of source rows: each dataset is packed up to its quota, so the number entered is exactly what runs, regardless of the dataset. The fixed dropdown becomes a checkable dataset library persisted in localStorage.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
