// Docs

Pooling

Use pooling-style inference for reward and progress scoring

Pooling inference is used for model APIs that return token logits, embeddings, or classification-like outputs instead of chat text.

For direct use, wrap a row function with generate_pooling:

async def score(row, generate_pooling):
    response = await generate_pooling(
        {
            "task": "token_classify",
            "messages": [{"role": "user", "content": "score this"}],
        }
    )
    return row.update({"raw_score": response})