Skip to content

Conversation

@shannonsands
Copy link
Contributor

PR Type

  • RL Environment PR - Complete Environment Snapshot & Zero-Training sections
  • Non-Environment PR - Complete Description, Related Issues & Type of Change sections

📝 General Information

Description

Integration & registry code for TextArena. Random sampling over all TextArena games. Minimal agent implementation, just parallel rollouts with max steps. Single agent trainer only, no full self play yet (opponents don't train)

🔖 Environment Snapshot

Field Your Entry
Environment Name TextArenaEnvMinimal
Short Description Train a policy on a vast number of different games
Category RLVR
Dataset Needed? No
External Deps textarena
Environmental Variables OPENROUTER_API_KEY (for some games like DebateEnv)
Compute Footprint Estimate Single node works

🧪 Zero-Training Test Results

Details

W&B Link:
https://wandb.ai/nous_research/hermes-qwen3-14b-grpo-textarena/runs/myzy6tx0

Examples of the Environment scoring a good example and a bad example:


✅ Developer & Reviewer Checklist

  • Code follows project style (black, isort, flake8 pass with pre-commit)
  • I have performed a self-review of my own code
  • I have commented my code, particularly in hard-to-understand areas
  • I have made corresponding changes to the documentation
  • My changes generate no new warnings
  • New and existing unit tests pass locally with my changes
  • Docstrings added for all new public classes / functions
  • If .env vars required, did you add it to the .env.example in repo root?

@shannonsands shannonsands changed the title WIP: Textarena minimal Textarena minimal Aug 12, 2025
@shannonsands shannonsands marked this pull request as ready for review August 12, 2025 04:45
@shannonsands shannonsands requested a review from dmahan93 August 12, 2025 04:45
f"[Traj {trajectory_idx}] Step {step_count} | Player {current_player} | Obs: {_obs_preview}..."
)

if current_player == self.config.training_player_index:
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

if it's multiplayer and the other players are the same policy why are we throwing them away

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

because I didn't want to overcomplicate things with this PR basically. We'd need to do group rollouts with all players & build their own ScoredDataGroups

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

we're basically doing that already minus building independent ScoredDataGroups

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok, can add it. But the other players trajectories won't all be starting from the same state (since player 1's moves can be totally different), so not sure how meaningful the advantage calculation for them will be

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

done

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

sorry I don't see that?

f"[Traj {trajectory_idx}] Step {step_count} | Player {current_player} | Obs: {_obs_preview}..."
)

if current_player == self.config.training_player_index:
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

sorry I don't see that?

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.

3 participants