Unverified50% confidenceOpinionExact time
ZY-S1's MATH-500 dropping from 4 to 0 is a typical manifestation of catastrophic forgetting caused by a distribution shift in the fine-tuning data.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/13/2026
First Seen
Valid until: 1/11/2027
Sources
Related Entities
Related Claims
UnverifiedPrice data typically exhibits a right-skewed distribution, which can cause Z-score methods based on standard deviation to miss low-end outliers.62% similarUnverifiedZ-score 标准化公式为 (x - mean) / std，将数据变换为均值0、标准差1的分布62% similarUnverifiedThe Z-score method flags data points that deviate from the mean by more than 2-3 standard deviations as outliers.61% similarUnverifiedTraditional RAG systems often suffer a significant drop in retrieval accuracy when faced with heterogeneous data.60% similarUnverifiedA related paper published in Nature in 2024 showed that repeatedly fine-tuning a model on its own outputs systematically causes the loss of tail information from the original data distribution.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1014228API
curl https://kongchang.com/api/v1/knowledge/claims/1014228MCP
get_claim(id=1014228)