Simple starter code for experiments on open-source LLMs. Built for my SPAR project participants, but anyone is welcome to use it.
# optional: create a virtual environment
python3 -m venv venv
source venv/bin/activate
# run from the root of the repo, this will install everything you need
pip install -e .To download Llama models from huggingface and/or use Claude API, add a .env file in the root of the repo with your API keys (see .env.example).
All code is in lmexp/
Example data and generation scripts using Claude API.
Example Llama 3 fine-tuning implementation. Quantizes to 8-bit. You may also want to try LoRA / PEFT methods / torchtune. Meta's fine-tuning example code can be found here.
Implementation of model-internals techniques like CAA and linear probing in terms of an abstract HookedModel class. An extended class, SteerableModel, is also provided for techniques that require modifying the model's activations.
See models/implementations/gpt2small.py for an example of how to use this class. The idea is that we can write a single implementation of a technique, and then apply it to any model we want. Note that this is very similar to the TransformerLens paradigm but pared down a lot to just provide the functionality we're likely to use. Feel free to use TransformerLens if you want more features.
Model implementations. Currently has:
- Gemma 2
- Llama 3.1
- Qwen 1.5
- GPT2 (useful for testing locally)
Jupyter notebooks demonstrating basic use-cases.
- Integrate with Gemma 2 SAEs / SAE feature steering
- Port over all the experiments / plotting code from CAA repo
- More contrast pair datasets