A library for finetuning LLMs (particularly the Llama family) and validating on AgentBench in one go.

Setup

• Initialize the submodules

git submodule init
git submodule update

• Install the requirements

conda create -n venv python=3.10
conda activate venv
pip install -r requirements.txt

# To evaluate on AgentBench
pip install -r evaluator/agentbench/AgentBench/requirements.txt
pip install -r evaluator/agentbench/AgentBench/src/tasks/knowledgegraph/requirements.txt
pip install -r evaluator/agentbench/AgentBench/src/tasks/lateralthinkingpuzzle/requirements.txt

• Finetune with QLoRA quantization

export HUGGINGFACE_API_TOKEN=<API_TOKEN>
python llama2_qlora.py configs/<filename>.yaml