iHore/J-Wash
8
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Support fitting lenses on any dataset and equal-parts mixes

Browse Source 
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>
This commit is contained in:
Extraltodeus
2026-07-14 05:08:52 +02:00
co-authored by Claude Opus 4.8
parent daeb127651
commit d9773394e3
4 changed files with 180 additions and 77 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+5 -3
View File
@@ -211,9 +211,11 @@ Gemma — the error shows up in the UI if so.)
### 6. Fit your own lens
In **Fit** (model unloaded, VRAM free), fit a lens on streamed WikiText across one
or more GPUs, with per-prompt checkpoints (stop/resume without loss) and weighted
merging. Metadata is written to `lenses/<name>/meta.json`.
In **Fit** (model unloaded, VRAM free), fit a lens across one or more GPUs on the
corpus of your choice: tick any HuggingFace dataset by id (WikiText by default),
or tick several to fit on an equal-parts mix. Per-prompt checkpoints give
stop/resume without loss, and multi-GPU slices are merged by weighted average.
Metadata is written to `lenses/<name>/meta.json`.
### CLI (no UI)