Unverified50% confidenceFactExact time
The third round of fine-tuning reduced the allow proportion in the training set to 61%, achieved by removing 242 safe samples already covered by deterministic rules, duplicating review/block examples, and adding 100 targeted examples.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/3/2026
First Seen
Valid until: 2/1/2027
Sources
Related Entities
Related Claims
UnverifiedThe retraining involved a clear trade-off: slight decreases in some very high scores on cleaner validation distributions, in exchange for improved real-world performance.71% similarUnverifiedThe encoder-only repair approach can reduce trainable parameters by 3× and optimizer-state memory by 3×.69% similarUnverifiedPerturbation-augmented training can only narrow the probe performance gap to -3.7 percentage points, with limited effectiveness.69% similarUnverifiedPushing training steps extremely high, piling up sample counts, and ignoring the validation set are the three pitfalls beginners most often fall into.69% similarUnverifiedApproaches to mitigating the coverage problem include introducing diversity penalty terms at the decoding stage, upsampling training on Semantic IDs of long-tail items, and adding coverage constraints at the ranking stage, but these methods often involve trade-offs against immediate metrics such as click-through rate.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130043API
curl https://kongchang.com/api/v1/knowledge/claims/1130043MCP
get_claim(id=1130043)