Unverified50% confidenceEventExact time
A Reddit user fine-tuned ModernBERT with 138,000 financial news headlines, and the training loss kept decreasing but the test set AUC stayed between 0.49 and 0.51.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/12/2026
First Seen
Valid until: 1/10/2027
Sources
Related Entities
Related Claims
UnverifiedTraining with the same recipe using only license-clean articles resulted in an accuracy drop of 0.0123 (95% interval 0.0034 to 0.0207).62% similarUnverified消融实验显示，若继续训练主干通用知识掉10%、数学掉17.8%；若两塔共享权重则通用任务掉26%，其余指标全线崩溃59% similarUnverified当训练窗口、门控窗口和报告窗口未严格分离时，PR-AUC达到0.72到0.76，严格分离后骤降至约0.51接近随机水平，这是数据泄漏的表现59% similarUnverifiedA reader trained on 20 million tokens further reduced overall LM loss but caused degradation on math tasks, so the author chose 15 million tokens as the balance point.59% similarUnverified在 Terminal Bench 2.1 测试中，OpenCode 对比里 89 个任务的失败数从 29 降到 16，多解决 13 题，分数提升 14.61 个百分点59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1005465API
curl https://kongchang.com/api/v1/knowledge/claims/1005465MCP
get_claim(id=1005465)