SFT

本文对比RLHF与DPO两种方法，揭示大语言模型如何通过偏好学习提升帮助性，解析训练三阶段及技术局限性。

Fireworks AI 现在支持 Kimi 2.7 的完全可训练，提供低成本构建竞争模型的方案。

Fine-tuning on proprietary data is the most strategic AI advantage; prompts are easily copied, while models trained on private data are hard to replicate. OpenAI is restricting this path—companies must act now to retain SFT control.

Fireworks AI 平台正式支持智谱 GLM 5.1 模型，提供 SFT/DPO 微调能力、200K 超长上下文窗口，专为长周期智能体编程微调优化，RL 训练即将上线。

Fireworks AI 宣布 Qwen 3.5 正式接入其训练平台，支持 SFT/DPO/RLHF 等全栈微调方式、256K 上下文及 LoRA/全参微调。

Fireworks AI 扩展其训练平台，支持 Nemotron 3 Ultra 模型的微调和 DPO 训练。

Spotify builds a highly steerable personalized recommendation system for 750M users and 100M+ tracks by transforming user action sequences into vectors and then tokens, combining content representation with LLMs.

Cursor团队在v9模型训练中贡献工程进展，但补充数据效果有限。

This article discusses the underlying mechanism of large model SFT (Self-Training), aiming to put an end to practice controversies and avoid fruitless computing power. By deeply understanding the SFT mechanism, engineers can more effectively utilize resources and avoid unnecessary calculations.