ContextQMD
Back to Sglang

Reward models

docs_new/docs/supported-models/reward_models.mdx

0.5.112.2 KB
Original Source

These models output a scalar reward score or classification result, often used in reinforcement learning or content moderation tasks.

They are executed with --is-embedding and some may require --trust-remote-code.

Example launch Command

<CodeGroup> ```shell Command python3 -m sglang.launch_server \ --model-path Qwen/Qwen2.5-Math-RM-72B \ # example HF/local path --is-embedding \ --host 0.0.0.0 \ --tp-size=4 \ # set for tensor parallelism --port 30000 \ ``` </CodeGroup>

Supported models

Model Family (Reward)Example HuggingFace IdentifierDescription
Llama (3.1 Reward / LlamaForSequenceClassification)Skywork/Skywork-Reward-Llama-3.1-8B-v0.2Reward model (preference classifier) based on Llama 3.1 (8B) for scoring and ranking responses for RLHF.
Gemma 2 (27B Reward / Gemma2ForSequenceClassification)Skywork/Skywork-Reward-Gemma-2-27B-v0.2Derived from Gemma‑2 (27B), this model provides human preference scoring for RLHF and multilingual tasks.
InternLM 2 (Reward / InternLM2ForRewardMode)internlm/internlm2-7b-rewardInternLM 2 (7B)–based reward model used in alignment pipelines to guide outputs toward preferred behavior.
Qwen2.5 (Reward - Math / Qwen2ForRewardModel)Qwen/Qwen2.5-Math-RM-72BA 72B math-specialized RLHF reward model from the Qwen2.5 series, tuned for evaluating and refining responses.
Qwen2.5 (Reward - Sequence / Qwen2ForSequenceClassification)jason9693/Qwen2.5-1.5B-apeachA smaller Qwen2.5 variant used for sequence classification, offering an alternative RLHF scoring mechanism.