Unverified50% confidenceFactExact time
QLoRA论文提出的NF4针对预训练权重服从正态分布的特性设计了最优量化格点
1
Sources
50%
Confidence
Long-term
Relevance
7/7/2026
First Seen
Sources
Related Claims
VerifiedQLoRA quantizes the base model to 4-bit precision for storage, while applying LoRA adapters on top of the quantized model for training.76% similarUnverifiedQuantization-Aware Training (QAT) lets the model retain a full-precision copy of weights, practicing rounding operations in each forward pass to learn weight distributions under extreme compression.74% similarUnverifiedQwen3采用四阶段训练流程，在后期专门针对混合推理能力进行对齐微调69% similarUnverifiedThe Scaling Law paper by Kaplan et al. in 2020 revealed that, under a fixed compute budget, there is an optimal ratio between model parameter count and the number of training tokens (the Chinchilla optimal).67% similarUnverified后训练的最优解是参考分布的指数倾斜，即最优分布密度正比于参考模型密度乘以奖励的指数：π*(x) ∝ π_ref(x) · exp(r(x)/β)65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/187835API
curl https://kongchang.com/api/v1/knowledge/claims/187835MCP
get_claim(id=187835)