Skip to content

Make the per-option token cap configurable (cfg option_max_len) - #38

Open
r33drichards wants to merge 2 commits into
mainfrom
claude/busy-keller-knhbhh
Open

r33drichards wants to merge 2 commits into
mainfrom
claude/busy-keller-knhbhh

Conversation

@r33drichards

Copy link
Copy Markdown
Owner

Options were cut to a hard-coded 48 tokens before the head_max_len split.
The cap is now cfg["option_max_len"] (default 48, unchanged), so a caller
can keep long option descriptions whole:

laya.load_vlm("thaitea/laya-vision", option_max_len=256, head_max_len=1024)

Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_011g7NrLwbcbKyqSPYGFgbop

Options were cut to a hard-coded 48 tokens before the head_max_len split.
The cap is now cfg["option_max_len"] (default 48, unchanged), so a caller
can keep long option descriptions whole:

    laya.load_vlm("thaitea/laya-vision", option_max_len=256, head_max_len=1024)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011g7NrLwbcbKyqSPYGFgbop
… taken answer with a per-row advantage)

A dataset record with an "advantage" is a policy-gradient row: its label is the answer the policy sampled
when it acted, and train() gives it -advantage * log p(answer) (weight w_pg) instead of the supervised
objective; batches without such rows train exactly as before. Used by jev-drone's RL loop, whose rollouts
score each decision with the simulator (group-relative advantages).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011g7NrLwbcbKyqSPYGFgbop

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants