Unverified50% confidenceFactExact time
当教师模型与学生模型使用不同分词器时，教师模型高达约85%的信息可能在token映射过程中丢失
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
跨分词器知识蒸馏：85%信息丢失的根因与链式法则解法
Related Claims
UnverifiedPerplexity measures a language model's ability to predict test text; the lower the value, the less surprised the model is by the text.64% similarUnverifiedPerez et al.的《Discovering Language Model Behaviors with Model-Written Evaluations》证实经过RLHF训练的模型普遍存在谄媚倾向63% similarUnverified主流扩散模型的文字生成依赖CLIP等视觉-语言预训练模型，其训练数据以英文为主，导致中文字符生成困难62% similarUnverifiedEnglish content typically accounts for 60%–80% of the training data in mainstream large language models.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/508976API
curl https://kongchang.com/api/v1/knowledge/claims/508976MCP
get_claim(id=508976)