Unverified50% confidenceFactExact time
In the first round of training data, 86% of rows were labeled allow, and the median length of real allow commands was 287 characters, whereas hand-written risky commands were usually only 33 to 40 characters, causing the model to learn the spurious signal that "short command = dangerous".
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedBefore testing, the semantic threshold was fixed at 0.81, a value calibrated on the Banking77 training set to guarantee that at least 98% of semantic hits point to the same intent; but on the actual test set the intent agreement rate of semantic hits was 96.7% to 96.8%, with each model having 30 to 31 hits on the wrong intent.71% similarUnverifiedThe 29-million-parameter encoder trained by the researchers achieved a popularity-weighted cross-fold AUC of only 0.486 on 985 statements, indistinguishable from the corpus prior.70% similarUnverifiedThe total of 2167 multiple-choice questions and 999 experimental records is still insufficient for training a large-scale generative model, and is better suited as an evaluation benchmark or fine-tuning seed.68% similarUnverified在小样本LoRA训练中，rank值设置存在偏差-方差权衡，实践中rank=4至32是常见的安全区间68% similarUnverifiedAfter 15 training epochs on the demonstration dataset, the text classification model achieves an accuracy of 99%, with a total of 110,000 model parameters (16,000 for Attention, 33,000 for FFN, and the remainder for Embedding)67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130041API
curl https://kongchang.com/api/v1/knowledge/claims/1130041MCP
get_claim(id=1130041)