Qwen3-1.7B-Wordle-RL
This model continues from Qwen3-1.7B-Wordle-SFT with reinforcement learning in the Wordle environment.
Research release
What the release provides
The model card identifies a separately trained reinforcement-learning checkpoint. The published recipe and environment are available in prime-rl.Qwen3-1.7B-Wordle-RL official model cardPrime RL Wordle training example
How to use it well
Practical guidance. Use held-out target words when comparing policies. Record completed games and illegal guesses separately so partial format rewards do not look like wins.
The published run uses complete game interactions
The example describes 100 reinforcement-learning steps with batches of 512 rollouts and a 4,096-token training context. Its evaluation uses target words outside the training set.Prime RL Wordle training example
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Repository license | Apache 2.0Qwen3-1.7B-Wordle-RL official model card |
Related models
- Qwen3-1.7B-Wordle-SFT
Prime Intellect ยท Prime RL research models
A supervised Qwen3 fine-tune that learns the response format for Wordle.
Launch date not confirmedResearch releaseDownloadable