| { | |
| "base_model": "Qwen/Qwen3-8B", | |
| "encoder_pooling": "last-token", | |
| "embedding_dim": 4096, | |
| "embedding_normalisation": "l2, applied client-side by clm.Embedder", | |
| "max_context_tokens": 2048, | |
| "truncation": "head (first N tokens), matching the vLLM serving path", | |
| "logit_scale": 100.0, | |
| "logit_scale_note": "exp(clm cfg logit_scale) = 100.81, clamped to 100.0 by HeadPair._load (heads.py:95). The head amplifies cosine error 100x.", | |
| "heads_repo": "Contrastive-LM/CLM-v0.1-8B", | |
| "heads_file": "CLM_v0.1-8B.pt", | |
| "note": "The projection heads are NOT quantised and are NOT re-shipped here. They live in the parent repo and clm-serve fetches them into ~/.cache/clm/.", | |
| "package": "clm_mlx", | |
| "mlx_lm_version": "0.31.3", | |
| "entrypoint": "python -m clm_mlx.server --model <dir> --port 8092" | |
| } | |