Free guide · Oct 11, 2026
Define the state, supply choices and inspect the ranking
How to define a state, write useful candidate actions, and inspect the ranking.
The short version
The released CLM v0.1 8B setup uses separate scoring heads on a frozen Qwen3-8B encoder. It ranks supplied candidates rather than generating text. The model card says its probabilities are relative to the candidate set, heads require compatible encoder embeddings and reported verifier benchmark results use fine-tuned heads. Weights and base encoder are identified as Apache 2.0. Official model card, code
What's in the full guide
- 011. Choose a small original routing problem
- 022. Make the candidate set fair
- 033. Inspect setup compatibility before inference
- 044. Submit the exact state and choices
- 055. Change one fact and compare
- 066. Keep model claims and experiment claims separate