Unverified50% confidenceFactExact time
Output distillation involves directly using text generated by a teacher model—such as dialogues and reasoning chains—as training data to fine-tune a student model, without requiring access to the teacher's internal parameters.
1
Sources
50%
Confidence
Long-term
Relevance
9/28/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn offline distillation, training data is generated once by the teacher model and fixed, with the student model learning from a static dataset.79% similarUnverified合成数据通常由一个强大的教师模型生成问答对、推理链或代码样本，再用于训练学生模型75% similarVerifiedHard distillation refers to the student model directly learning the final output labels or generated text of the teacher model, while soft distillation requires the student model to learn the complete probability distribution (logits) output by the teacher model.75% similarUnverified模型蒸馏是指用大型教师模型的输出训练更小的学生模型，仅需调用API获取输出，不涉及权重复制，其法律边界模糊，全球尚无先例性司法判决明确定性74% similarUnverifiedModel Distillation in machine learning refers to training a small student model using the outputs of a large teacher model, in order to compress model size while retaining capability.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/956807API
curl https://kongchang.com/api/v1/knowledge/claims/956807MCP
get_claim(id=956807)