Contrastive-LM/CLM v0.1 8B

• License: Apache-2.0
• Release date: Not documented; 2026-09-21 is the first HF publication (repository creation), not a verified release date.
• Modalities: Input: text; Output: embeddings
• Original model: Contrastive-LM/CLM-v0.1-8B

Description

CLM learns to match states with candidate actions using contrastive training. The original model places separate state and action projection heads on a frozen Qwen3-8B encoder. This registry entry serves only that encoder's last-token pooled embeddings, not the original model's decision or ranking head stack.

Run

self serve clm-v0.1:8b

Here is a link to the API docs: Embeddings API.

Example Request

With the server running, open another terminal and send a request:

curl http://localhost:8080/v1/embeddings \
	-H 'Content-Type: application/json' \
	-d '{"model":"clm-v0.1:8b","input":"A customer was charged twice for an order."}'

The server returns JSON containing the text embedding.

Engine Parameters

Served by upstream llama-server through the embedding adapter. These are the registry-configured startup defaults; text-generation sampling settings do not control embeddings.

Parameter	Default	What it controls
context_size	8192	Total llama-server token context, shared across parallel slots; accepts 512 to 1048576.
pooling	last	Embedding pooling; the CLM heads require last. Adapter accepts none, mean, cls, last, rank, but these are not interchangeable for CLM.

Override startup settings with --set <parameter>=<value>.

Quantizations and Files

Default variant: q4. Sizes and SHA-256 hashes are pinned in the model registry.

q4 (default)

File	Role	Size	SHA-256	Origin
Qwen3-8B-Q4_K_M.gguf	model	5.11 GiB (5489669760 bytes)	373e464e5310bda8fc54c1b67b36c2b860be5f03e5e0d3d046f5ef8da6b7e7d6	czl/CLM-v0.1-8B-GGUF

Total download: 5.11 GiB (5489669760 bytes).

Notes

• English-focused; intended for embedding and ranking workflows, not text generation. The registry advertises 32768 tokens upstream, but configures an 8192-token total server context.
• The original projection heads are not downloaded or served. Typed decisions and candidate probabilities require the separate upstream CLM head stack; this server exposes /v1/embeddings only.
• The GGUF author warns that Q4_K_M is not recommended for ranking because quantization changes decisions. Validate on your own data; only q4 is currently registered here.
• Upstream verifier benchmark results use fine-tuned heads, not this encoder alone or the base head checkpoint zero-shot.
• Download size is not the total runtime memory requirement. Origin links use main because this registry entry pins a hash, not a repository revision.

Benchmarks

No submitted benchmark reports yet.