SFT 的似然目标画不出相对偏好，只能到一个由数据规模决定的天花板。RLHF 用奖励模型把偏好压成标量再接 PPO，代价是 KL 约束、采样成本…

I’m afraid my off the cuff response the other day wasn’t too well though…

Today I’d like to discuss a beautiful inequality in graph theory, namely…

预训练模型会补全，不等于会按人类意图回答。本文把 instruction tuning 的逻辑讲清楚：SFT 数据从哪里来，prompt-res…

Let’s write a malloc and see how it works with existing programs!

I have added another essay to my career advice page, inspired partly by …

全参数微调的代价 → Adapter / Prefix Tuning / Prompt Tuning 的早期尝试 → LoRA 低秩分解的核心洞…

People frequently 1 think that I’m very stupid. I don’t find this surpri…

Ben Green and I have just uploaded our paper “The quantitative behaviour…

这篇文章只比较文本模型里最重要的三类预训练目标：GPT 的自回归语言建模、BERT 的掩码语言建模、T5/BART 的去噪式序列到序列。重点不是…