feat(recipe): add TMax terminal-agent reproduction - #141
Draft
kylemontgomery1 wants to merge 3 commits into
Draft
Conversation
Provide a reproducible Qwen3.5-9B training recipe for terminal agents. Assisted-by: GPT-5.6 Sol Co-authored-by: Cursor <cursoragent@cursor.com>
Pin the recipe to verl PR #7458 and opt out of deferred synchronization so the fused untied LM head can train under FSDP2. Assisted-by: GPT-5.6 Sol Co-authored-by: Cursor <cursoragent@cursor.com>
Use the current agent-loop API so TMax can continue after tool calls instead of failing after the first turn. Co-authored-by: Cursor <cursoragent@cursor.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Background & Motivation
TMax is an RL recipe for training long-horizon terminal agents using the released TMax-15K environments. This PR adds an end-to-end reproduction of the paper’s main Qwen3.5-9B experiment to
verl-recipe. The motivation, design, and experimental setup were discussed in RFC #133.Key Changes
Add a new
tmax/recipe containing:REQUIRED_VERL.txtrecording the tested upstream revision.Validation
The recipe was validated with a full Qwen3.5-9B training run. The best checkpoint improved Terminal-Bench 2.1 from 21.5% to 29.8% under a 64k token budget and from 24.7% to 33.9% under the full 256k context window. Training curves and additional evaluation details are available in RFC #133.
Ruff lint and formatting checks pass.
Compatibility
tmax/; no verl core source is modified.