Unverified60% confidenceFactExact time
Mainstream code large models all adopt a two-stage strategy of "pretraining + instruction fine-tuning," with the second stage using alignment techniques such as RLHF or DPO.
2
Sources
60%
Confidence
Long-term
Relevance
7/6/2026
First Seen
Sources
Related Claims
VerifiedEach RL training step requires running inference (generating rollouts) and backpropagation (updating weights) simultaneously, making the compute requirements far higher than ordinary fine-tuning.76% similarUnverifiedRLHF的人类介入发生在训练阶段并固化为模型权重，属于隐式编码，而表征工程干预在推理时操控内部表征，具有即时性和可逆性75% similarUnverifiedIn ChatGPT's training, the pre-training phase uses unsupervised learning for next-token prediction, the instruction fine-tuning phase introduces supervised learning, and finally RLHF is used to align the model's outputs.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/114501API
curl https://kongchang.com/api/v1/knowledge/claims/114501MCP
get_claim(id=114501)