Contrastive-LM/CLM

选用指南

AI 根据 README 整理 · 2026/10/14 18:41:04

使用方式

通过 pip 安装后，使用命令行启动服务，并通过 Python 客户端或 HTTP API 进行调用。

查看 README 依据

We invite the community to plug it into their own agents and benchmarks! --- ## Installation ```bash pip install contrastive-lm ``` To install the latest from a clone: ```bash pip install -e . ``` --- ## Quickstart ### Serve ```bash # 1. encoder (Qwen3-8B embeddings) vllm serve Qwen/Qwen3-8B --served-model-name qwen3-8b --runner pooling --max-model-len 2048 --port 8090 & # 2. CLM API on :8700 (downloads the 75 MB reference head on first run) clm-serve ``` States longer than 2048 tokens are truncated. For longer states, raise both limits

together, e.g. `--max-model-len 8192` on `vllm serve` and `clm-serve --max-tokens 8192` (needs more GPU memory). ### Ask typed questions about a state ```python from clm import CLMClient, Choice, Noul, Score client = CLMClient() # CLM_BASE_URL (default http://127.0.0.1:8700), CLM_API_KEY r = client.system_one( state="Customer: my invoice was charged twice and nobody answers the phone!", questions={ "urgency": Noul(instructions="Is this urgent?"), "department": Choice(instructions="Which team should handle this?",

部署要求

需要 GPU 运行 vLLM 编码器（如 Qwen3-8B），并安装相关依赖。

查看 README 依据

We invite the community to plug it into their own agents and benchmarks! --- ## Installation ```bash pip install contrastive-lm ``` To install the latest from a clone: ```bash pip install -e . ``` --- ## Quickstart ### Serve ```bash # 1. encoder (Qwen3-8B embeddings) vllm serve Qwen/Qwen3-8B --served-model-name qwen3-8b --runner pooling --max-model-len 2048 --port 8090 & # 2. CLM API on :8700 (downloads the 75 MB reference head on first run) clm-serve ``` States longer than 2048 tokens are truncated. For longer states, raise both limits

├── evaluation/bon_eval.py # unified best-of-N evaluation ├── preprocessing/hf_embeddings.py # embedding dir <-> Hugging Face dataset ├── requirements.txt # pip install -r requirements.txt (clm + torch + vLLM + example deps) ├── examples/ # CLM vs Jev on the T-Rex runner (examples/t_rex/README.md) │ ├── common.py # one client for both endpoints: retries, latency, cache │ └── t_rex/ # Chrome dinosaur game in real time (run.py --model clm|jev) └── docs/FINETUNING.md # the fine-tuning guide ```

已知限制

状态超过 2048 tokens 会被截断，需提高限制并占用更多 GPU 内存。

查看 README 依据

We invite the community to plug it into their own agents and benchmarks! --- ## Installation ```bash pip install contrastive-lm ``` To install the latest from a clone: ```bash pip install -e . ``` --- ## Quickstart ### Serve ```bash # 1. encoder (Qwen3-8B embeddings) vllm serve Qwen/Qwen3-8B --served-model-name qwen3-8b --runner pooling --max-model-len 2048 --port 8090 & # 2. CLM API on :8700 (downloads the 75 MB reference head on first run) clm-serve ``` States longer than 2048 tokens are truncated. For longer states, raise both limits

together, e.g. `--max-model-len 8192` on `vllm serve` and `clm-serve --max-tokens 8192` (needs more GPU memory). ### Ask typed questions about a state ```python from clm import CLMClient, Choice, Noul, Score client = CLMClient() # CLM_BASE_URL (default http://127.0.0.1:8700), CLM_API_KEY r = client.system_one( state="Customer: my invoice was charged twice and nobody answers the phone!", questions={ "urgency": Noul(instructions="Is this urgent?"), "department": Choice(instructions="Which team should handle this?",