Fix VLM captioning label alignment and add an EOS stop target - #196

Merged

amazloumi merged 11 commits into

mainfrom

fix/vlm-next-token-labels

Sep 8, 2026

Commits on Sep 2, 2026

Fix - Shift VLM captioning labels for next-token loss
amazloumi
committed
c9696c8
Update video-dataset tests + docstrings for next-token labels
amazloumi
committed
48ee0ea
fix docstrings and changelog for next-token labels
amazloumi
committed
1f250b9
Append EOS to caption targets so captions learn to stop
amazloumi
committed
d59807b
Require vlm.max_text_len >= 2 (a caption token plus its target)
amazloumi
committed
849eaaa
Cover the no-EOS-tokenizer branch of _tokenize_and_mask
amazloumi
committed
c150564
Cross-check VLM labels against a derived next-token contract
amazloumi
committed
2aa755d
Note the objective change and the max_text_len floor in the changelog
amazloumi
committed
71171c1
Cover prompt lengths at and beyond max_text_len in the contract cross-check
amazloumi
committed
18c1234
Clarify that an over-long prompt is not rejected today
amazloumi
committed
02f3f4f
Run the label-contract cross-check on the no-EOS tokenizer branch too
amazloumi
committed
408ffc3