SkillCompiler/data/skills-bench/tasks/debug-trl-grpo/task.md

schema_version, metadata, verifier, agent, environment

schema_version	metadata	verifier	agent	environment
1.3	author_name author_email difficulty category subcategory category_confidence task_type modality interface skill_type tags Jiajun Bao jiajunb@alumni.cmu.edu hard software-engineering debugging high debugging repair source-code terminal python debugging-heuristic library-api-usage post-training trl debugging reasoning-model grpo	type timeout_sec service hardening test-script 120.0 main cleanup_conftests true	timeout_sec 3600.0	network_mode build_timeout_sec os cpus memory_mb storage_mb gpus public 2400.0 linux 4 2048 10240 0

debug-trl-grpo

I am training a countdown math task using TRL for my model with GRPO, but the new model shows no improvement. The source code is at /app/trl. Check if there is any bug and fix them.

DO NOT modify the training script (/app/train_grpo.py) or reward function (/app/reward_fn.py)