Key Features
State-action scoring.
Frozen Qwen3-8B backbone.
Separate projection heads.
Bidirectional InfoNCE training.
Candidate action ranking.
Verifier fine-tuning.
Reusable action embeddings.
A contrastive-lm Python package.
About CLM-v0.1-8B
Agent developers can reuse cached action embeddings and evaluate a faster decision or verification component alongside a generative model. Core capabilities include State-action scoring, Frozen Qwen3-8B backbone, Separate projection heads, and Bidirectional InfoNCE training.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Candidate action ranking, Verifier fine-tuning, Reusable action embeddings, and A contrastive-lm Python package.