Unverified50% confidenceFactExact time
The loss during training is calculated per token, and long samples contribute far more gradient than short samples, so a 10% share of rows can absolutely correspond to more than 50% of the token volume and gradient signal.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe power-law exponent between test loss and compute is approximately 0.05, meaning that every 10x increase in compute yields approximately a 12% improvement in model performance.76% similarUnverifiedWhen sampling eight candidates without constraints from the model, 10.2% of inputs cannot produce any valid output under limited training data, and this proportion drops to 0.4% when using the full training data73% similarUnverifiedIn CLM training, pretraining alone scores 52% on the hard negative sample test; adding the hard negative sample stage jumps it to 69%; training with hard negative samples from day one peaks at only 62% and overfits quickly; omitting replay causes the hard negative sample score to drop from 69% back to 56%.71% similarUnverifiedIn zero-shot testing on an unseen grid of 6,470 buses (approximately 40% larger than any grid in training), the cost error increases by about 10%, but the model can still rank scenarios by cost with near-perfect accuracy.71% similarUnverifiedFine-tuning on a few hundred real data samples on top of synthetic pre-training can often improve mAP by 10 to 20 percentage points.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1018230API
curl https://kongchang.com/api/v1/knowledge/claims/1018230MCP
get_claim(id=1018230)