Unverified50% confidenceBenchmarkExact time
The KLD of models uncensored via traditional full fine-tuning typically falls in the range of 0.1 to 0.5 or even higher, while LoRA-based methods generally fall between 0.05 and 0.15.
1
Sources
50%
Confidence
Long-term
Relevance
9/7/2026
First Seen
Sources
Related Entities
Related Claims
Unverifiedk=5时可以以95%置信度区分成功率0.9和0.7的模型，k=10时能识别更细微的差异68% similarUnverifiedLoRA由微软研究院于2021年提出，对权重更新量ΔW进行低秩分解ΔW = A·B，参数量通常可压缩至原来的1/100以下63% similarUnverifiedLoRA decomposes the fine-tuning weight change ΔW into the product of two low-rank matrices, with a parameter count of only 0.1%–1% of the original model, achieving a parameter compression ratio as high as 256× when r=8 and d=4096.62% similarUnverifiedThe GLM-OCR model has 0.9B (900 million) parameters.62% similarUnverifiedGAE通过引入λ参数在偏差和方差之间折中，λ=1退化为蒙特卡洛估计，λ=0退化为单步TD估计，实践中λ=0.95左右通常能取得良好平衡61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/871912API
curl https://kongchang.com/api/v1/knowledge/claims/871912MCP
get_claim(id=871912)