wordle-grpo-Qwen3-1.7B

Fine-tune of Qwen/Qwen3-1.7B trained to play Wordle using GRPO via TRL and the OpenEnv framework.

Training details

• Method: GRPO (Group Relative Policy Optimization)
• Base model: Qwen/Qwen3-1.7B (thinking disabled at inference — enable_thinking=False)
• Environment: TextArena Wordle-v0 via OpenEnv, hosted on a HuggingFace Space
• Compute: NVIDIA A100-80GB on Modal (~94 minutes)
• Steps: 62 gradient updates, 4 rollouts per prompt
• Total games played: ~15,800 live Wordle episodes during training

Results (50-episode eval)

	Baseline (Qwen3-1.7B)	GRPO fine-tuned
Win rate	0%	8%
Avg shaped reward	1.33	1.44
Avg yellow reward	0.184	0.236

Visualize in Trackio