🤖

GRPO CodeReviewEnv

Reinforcement Learning · Bug-Fix Agent · Auto Difficulty Escalation

Qwen2.5-Coder-32B HF Router GRPO Training Exec + LLM Judge Rewards
🎯
Current Level
HARD
📊
Total Episodes
10
🔥
Win Streak
0
⚡
Last Reward
0.830

📋 Training Stats
Extreme
3
0.938
0.83
0.69
✅ Mastered

📡 Live Episode Feed (last 20)
10
Medium
0.830
████████░░
BugGenerator LLM → buggy challenge easy/medium/hard/extreme challenge RL Agent Qwen2.5-Coder-32B via HF Router agent_fix(challenge) fixed code HFRewardEvaluator ExecScore 60% — run tests JudgeScore 40% — LLM eval Final = 0.6×exec + 0.4×judge reward DifficultyEscalator rolling avg (window=5) win/lose streaks escalate / stay / drop → progress.json next difficulty level GRPO CodeReviewEnv — Training Architecture LLM Agent · Execution Reward · LLM-as-Judge · Auto Difficulty Escalation
5 100

Auto-refreshes every 3s | Escalate threshold: 0.8 | Window: 5