Unverified50% confidenceFactExact time
In self-training, the training temperature affects the sampling distribution during data generation: temperature T adjusts the model's output probabilities via p_i ∝ exp(z_i/T), with T>1 making the distribution flatter and T<1 making it sharper.
1
Sources
50%
Confidence
Long-term
Relevance
11/9/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe complete training objective is L = α·L_CE(y, p_s) + (1-α)·T²·L_KL(p_t^T, p_s^T), where T² is a gradient correction factor introduced by temperature scaling.77% similarVerifiedIn the standard split conformal prediction pipeline, the model is first trained on a training set, then nonconformity scores are computed for each sample on an independent calibration set, and the threshold is taken as a specific quantile of these scores.64% similarVerifiedWhen parameter counts are similar, differences in training data quality and distribution often determine the performance ceiling in specific scenarios more than architectural differences.64% similarUnverifiedThe overall training process first calibrates the relative relationship between label scores and feature scores, then increases the attention scale at a rate of t^(1/4), causing the overall threshold error to converge at a rate of O(t^(-1/4))63% similarUnverifiedFor DPO training pipelines that use a learning rate scheduler, the peak learning rate and warmup steps also need to be adjusted accordingly with model scale.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1149952API
curl https://kongchang.com/api/v1/knowledge/claims/1149952MCP
get_claim(id=1149952)