Interest in a verl-based training pipeline alongside SkyRL?

Ouverte
#80 5 commentaires 0 réactions 0 personnes assignées Voir sur GitHub

Personne n'a encore pris cette issue.

Évaluation

Difficulté
5/5
Temps estimé
Plus d'une semaine
Accessibilité débutants
20/100
Type d'issue
Fonctionnalité
Clarté
Plutôt claire
Activité
Active
Stack technique
python

Piste de recherche

The issue describes a proposed verl training path but names no repository files or tests to start from. First read the existing SkyRL pipeline and ask maintainers whether they want this work and which scope they prefer; the thread has not recorded a decision. Done means an agreed scope and an accepted implementation, with the synchronous-trainer limitation documented.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

Summary

We completed a full port of the CodeScout agentic RL training pipeline (issue → file localization) from SkyRL to verl (v0.9.1), preserving task, reward, and evaluation semantics. Before investing in a large PR, we'd like to check whether you're interested in accepting a verl-based training path alongside the current SkyRL pipeline.

What the port includes

  • A verl agent-loop adapter for the CodeScout multi-turn rollout: terminal search tools, localization_finish, prompt/tool-schema plumbing, and workspace isolation, wired into verl's multi-turn rollout and training loop.
  • Reward bridging to multilevel_localization_f1_reward (file/module/entity F1, sum over 0–3), with the private-label boundary kept out of the actor's visible inputs.
  • ~3,300 lines of tests, including:
    • loss/gradient/advantage numerical alignment against SkyRL fixed tensors (10 layered checks),
    • a 15-case reward contract suite,
    • adapter unit + CLI end-to-end tests.
  • Data prep / training / evaluation entry scripts and environment setup scripts.

Reproduction evidence

All numbers below are on the localization versions of SWE-bench Verified (500 tasks) / Lite / Pro, sum-F1 macro average on a 0–3 scale, starting from Qwen/Qwen3-4B-Instruct-2507 with GRPO advantage + GSPO policy loss, 200 updates on 8×H100.

  • Under the official 15-turn protocol with official ground truth, our verl-trained reproduction reaches or slightly exceeds the released official CodeScout-4B reference model under the same pipeline (same-pipeline comparison against the released model, not the paper tables):
    • Verified: 0.815 / 0.783 (two rollout seeds) vs. official 0.730
    • Lite: 0.636 / 0.552 vs. official 0.497
    • Pro: 0.084 / 0.078 vs. official 0.070
  • Under our frozen internal protocol, the untrained base scores 0.262 → 0.762 / 0.733 after training (mean paired difference +0.486, task-paired bootstrap 95% CI [0.407, 0.565]).

Known deviation from the original pipeline

Our current trainer is synchronous (staleness = 0), whereas the original SkyRL pipeline is fully async. We treat this as a documented limitation rather than a silent change; migrating to verl's async trainer would require redoing a matched baseline.

Questions for maintainers

  1. Would you be open to a PR adding a verl training path (e.g., under an opt-in directory or flag, keeping SkyRL as the default)?
  2. If yes, what scope would you prefer?
    • (a) adapter + tests only,
    • (b) adapter + tests + entry scripts + user-facing docs,
    • (c) something narrower — we're happy to split into smaller incremental PRs.
  3. Any requirements we should know about (license headers, supported verl versions, CI)?

Happy to provide per-task audit data or the full experiment records if useful.

Langage dominant
Python
Étoiles
93
Forks
16
Merge moyen
9 h 57 min
PR mergées (30 j)
1

Préparer son environnement

Ce projet ne fournit ni conteneur de développement, ni Dockerfile, ni guide de contribution : l'installation est à votre charge. Commencez par son README, et consultez notre guide de la première contribution pour les étapes générales.

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de OpenHands/codescout

Toutes les issues de OpenHands/codescout

Issues similaires

Plus d'issues Python

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.