llm-tldr's Introduction

llm-tldr

Short description of LLM research and development

Papers

Chain of Thought (Jason Wei 2023)

Paper: Chain Of Thought

A new prompting technique to improve LLM reasoning performance.

Example 1:

Example 2:

Result: for Math, CoT can improve 2x solve rate

One liner command

curl -X POST "https://api.openai.com/v1/chat/completions" \
        -H "Authorization: Bearer YOUR_API_KEY" \
        -H "Content-Type: application/json" \
        -d '{
               "model": "gpt-3.5-turbo",
               "messages": [
                   {"role": "system", "content": "You are a helpful assistant."},
                   {"role": "user", "content": "Explain the steps to find the least common multiple (LCM) of these numbers: [12, 15, 18] and provide the answer in JSON format with your thoughts and the final answer."}
               ],
               "max_tokens": 250,
               "temperature": 0.3,
               "n": 1,
               "stop": null
            }'

[Github Workspace][https://github.com/codespaces/new/khangich/llm-tldr]

ReAct (Shunyu Yao 2023)

Paper: ReAct

Example 1:
Example 2:
Result
Weblink diagram

+-------------------+      +-------------------+      +-------------------+
|   Reset Env with  |      |   Generate Thought|      |   Execute Action  |
|   Question (idx)  |----->|   & Action using  |----->|   in Environment  |
|                   |      |   LLM Function    |      |   & Get Obs       |
+-------------------+      +-------------------+      +-------------------+
           |                        |                          |
           |                        |                          |
           v                        v                          v
+-------------------+      +-------------------+      +-------------------+
|   Update Prompt   |<-----|   Format Response |<-----|   Check if Done   |
|   with Question   |      |   & Handle Errors |      |   & Update Prompt |
+-------------------+      +-------------------+      +-------------------+

Recommend Projects