openenv/wordle-tool-calls-sft

Wordle tool-call SFT

1,000 winning Wordle games written as tool-calling conversations, for a short SFT warmup before training with GRPO on the OpenEnv Wordle environment. Used in the OpenEnv Wordle tutorial to train openenv/Qwen3-1.7B-Wordle-SFT.

A solver played every game against the TextArena Wordle-v0 environment from OpenEnv 0.10.0, run in-process. Each turn it keeps the words that fit all the feedback so far and picks the guess that splits them best.

• messages: the user prompt (the instructions plus the environment's opening text), then one assistant turn per guess, with a short reasoning_content and a call to the guess tool, each followed by a tool message with the environment feedback. The last assistant turn states the answer after the environment's win message.
• tools: the guess tool schema, as a JSON string.

This is the format TRL's GRPOTrainer renders with environment_factory, so a model trained on it plays the same way during RL. Train with SFTTrainer and assistant_only_loss=True.

The solver wins every game, in 2.7 guesses on average, and the reasoning is about 50 tokens per guess. Wordle-v0 draws the secret from 156 common nouns, so every one of them appears several times.

Regenerate

uv run generate_sft_data.py --games 1000 --push-to-hub <user>/wordle-tool-calls-sft

The script is in this repo and in OpenEnv examples/wordle.