https://github.com/hunkim/o
Toy O
https://github.com/hunkim/o
Last synced: 8 months ago
JSON representation
Toy O
- Host: GitHub
- URL: https://github.com/hunkim/o
- Owner: hunkim
- License: mit
- Created: 2024-09-14T20:18:41.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2024-09-21T08:27:13.000Z (over 1 year ago)
- Last Synced: 2025-03-31T05:04:43.870Z (10 months ago)
- Language: Python
- Size: 34.2 KB
- Stars: 16
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# o: Step-by-Step Reasoning Demos
This repository contains toy demos for step-by-step reasoning, inspired by:
- STaR paper https://arxiv.org/abs/2203.14465
- OpenAI's o1
- Reflection techniques
- SkunkworksAI/reasoning-0.01 dataset
- Implementation from [bklieger-groq/g1](https://github.com/bklieger-groq/g1)
- And many more
## Quick Demo
- toy-o1: https://toy-o1.streamlit.app/
- toy-o2: https://toy-o2.streamlit.app/
## Overview
We present two approaches:
1. **o1**: Fixed reasoning based on the SkunkworksAI/reasoning-0.01 dataset. You can also leverage web search.
2. **o2**: Relies on the LLM's basic planning skills (code mostly reused from bklieger-groq/g1).
Both approaches have their pros and cons, providing interesting comparisons.
We use [Solar-Pro Preview](https://huggingface.co/upstage/solar-pro-preview-instruct) as the base LLM, but you can try others using Langchain.
## Running Locally
1. Clone this repository
2. Add 'UPSTAGE_API_KEY' to your environment variables (use .streamlit config or .env)
3. Run `make o1` or `make o2`
## Changing Base LLMs
Replace the following code with your preferred LLM:
```python
langchain_upstage import ChatUpstage as Chat
llm = Chat(model="solar-pro")
```
## Limitations
Solar-Pro is only 22B size mode with 4K context windows. https://huggingface.co/upstage/solar-pro-preview-instruct
## Contributions
comments, pull request are always welcome