Unverified50% confidenceBenchmarkExact time
In the full-text screening stage, rationale statements written by researchers can improve model performance by approximately 15%
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified针对金融领域微调的嵌入模型（如FinBERT）相比通用嵌入模型在财报问答任务上准确率高出15%以上73% similarVerifiedResearch indicates that high-quality prompts can improve model output accuracy by 30%–50%.72% 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72% 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%.71% similarUnverified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70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/893423API
curl https://kongchang.com/api/v1/knowledge/claims/893423MCP
get_claim(id=893423)