Transformer和LLM前沿内容（3）：LLM Post-Training

视频链接：bilibili

文章目录

• • • [1. LLM Fine-Tuning](#1. LLM Fine-Tuning)
    • • [1.1 Supervised Fine-Tuning (SFT)](#1.1 Supervised Fine-Tuning (SFT))
      • [1.2 Reinforcement Learning from Human Feedback (RLHF)（重点）](#1.2 Reinforcement Learning from Human Feedback (RLHF)（重点）)
      • [1.3 Parameter Efficient Fine-Tuning (PEFT)（重点）](#1.3 Parameter Efficient Fine-Tuning (PEFT)（重点）)
      • • [1.3.1 BitFit](#1.3.1 BitFit)
        • [1.3.2 TinyTL: Lite Residual Learning](#1.3.2 TinyTL: Lite Residual Learning)
        • [1.3.3 Adapter](#1.3.3 Adapter)
    • [2. Multi-modal LLMs](#2. Multi-modal LLMs)
    • • [2.1 Cross-Attention Based: Flamingo](#2.1 Cross-Attention Based: Flamingo)
      • [2.2 Visual Tokens as Input: PaLM-E, VILA](#2.2 Visual Tokens as Input: PaLM-E, VILA)
      • [2.3 Enabling Visual Outputs: VILA-U](#2.3 Enabling Visual Outputs: VILA-U)
    • [3. Prompt Engineering](#3. Prompt Engineering)
    • • [3.1 In-Context Learning (ICL)](#3.1 In-Context Learning (ICL))
      • [3.2 Chain-of-Thought (CoT)](#3.2 Chain-of-Thought (CoT))
      • [3.3 Retrieval Augmented Generation (RAG)](#3.3 Retrieval Augmented Generation (RAG))

1. LLM Fine-Tuning

1.1 Supervised Fine-Tuning (SFT)

---

1.2 Reinforcement Learning from Human Feedback (RLHF)（重点）

---

1.3 Parameter Efficient Fine-Tuning (PEFT)（重点）

1.3.1 BitFit

1.3.2 TinyTL: Lite Residual Learning

1.3.3 Adapter

---

2. Multi-modal LLMs

2.1 Cross-Attention Based: Flamingo

---

2.2 Visual Tokens as Input: PaLM-E, VILA

---

2.3 Enabling Visual Outputs: VILA-U

---

3. Prompt Engineering

3.1 In-Context Learning (ICL)

---

3.2 Chain-of-Thought (CoT)

---

3.3 Retrieval Augmented Generation (RAG)