Unverified50% confidenceOpinionExact time

少量高质量的偏好数据往往比大量低质量数据能产生更好的对齐效果

1

Sources

50%

Confidence

Long-term

Relevance

8/6/2026

First Seen

Sources

Code Preference Dataset: Manually Reviewed Training Material for DPO/RLHF Fine-Tuning

Related Claims

Unverified标签粒度越细触达相关性越高，用户负反馈率越低，账户质量评分也随之受益72% similar Unverified第一方数据比第三方购买的数据价值更高，因为它真实、及时且专属70% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/696962

API

curl https://kongchang.com/api/v1/knowledge/claims/696962

MCP

get_claim(id=696962)