Unverified50% confidenceFactExact time
On synthetic biography data, when validation loss differs by only 0.5% to 2%, NVFP4 requires 1.32 to 1.52 times as many exposure counts as BF16 to reach the same rare-fact recall rate.
1
Sources
50%
Confidence
Long-term
Relevance
11/10/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedUnder overexposed conditions, recall stays at 1.000, but precision drops to 0.933, with 15 extra false positives.67% similarUnverifiedPrecision loss comparison by quantization level: Q5 is nearly lossless, Q4 has approximately 2%–5% loss, Q3 has approximately 8%–15% loss, and Q2 has approximately 30%–50% loss.67% similarUnverifiedThe beta value of the F2 score is set to 2, meaning recall is weighted 4 times more than precision66% similarUnverifiedFine-tuned DistilBERT achieves F1=0.9993 on in-distribution testing, but loses 12.9 mean average precision points under topic shift, a greater degradation than the linear model's 5.2 points.66% similarUnverified若单一模型漏检率为 p,两个训练背景不同的模型同时漏检同一错误的概率近似降至 p²,20%漏检率理论上可降至 4%66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1154468API
curl https://kongchang.com/api/v1/knowledge/claims/1154468MCP
get_claim(id=1154468)