Unverified50% confidenceSolutionExact time
靶向数据工程中，针对性数据占总微调数据的 5%-15% 通常能取得良好效果，过高比例可能导致模型在其他维度退化
1
Sources
50%
Confidence
Long-term
Relevance
8/25/2026
First Seen
Sources
Related Entities
Related Claims
Partially VerifiedThe quality of fine-tuning data is far more important than quantity; typically, a few hundred to a few thousand high-quality annotated data samples can significantly improve a model's performance on a specific task.75% similarUnverifiedScale AI的研究团队发现，部分模型在被污染的基准上的表现比在新编写的同等难度题目上高出10-15个百分点74% similarUnverifiedResearch shows that on certain open-source benchmarks, the n-gram overlap rate between the training data of leading large models and the evaluation data can be as high as 15%–30%.72% similarUnverifiedFor small models with limited parameter counts, data quality can have an even greater impact on final performance than data quantity.71% similarUnverified研究表明，即使在 GSM8K 或 MATH 数据集上接近满分的模型，面对稍作变形或需要跨领域组合的问题时性能下降可达30-50个百分点70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/798533API
curl https://kongchang.com/api/v1/knowledge/claims/798533MCP
get_claim(id=798533)