README.md

Task-specific Tool Retrieval Dataset Generation

To generate your own task-specific tool-retrieval, follow these 3 steps:

1. Prepare your tools

Prepare a file containing your tools and descriptions with the following format

{
    "<your function name>": {
        "new_func": "<new name>", # We used this field in some ToolBank datasets to assign fake names during dataset generation but it is completely optional. You can just use <your function name>,
        "description": "<function description>"
    },
    ...
}

2. (Optional) Write 2-3 seed in-context examples (ICE)

You can optionally write some initial ICEs for your task-specific dataset. For this, you can create task-specific InstructionGenerator class that inherits from BaseInstructionGeneratorand a InstructionPolisher class that inherits from BaseInstructionPolisher. You can check some examples of this in /bank_aws, /bank_numpy, and /bank_pandas folders. After creating the classes and the ICE according to the appropriate format, add these to main.py

3. Generate data

Use the generate_data_script.sh to generate synthetic data. You need to specify which model to use and appropriate file paths.

DATASET="..." # (numpy/pandas/aws)
MODEL="..."
TOOLS_PATH="..."
SAVE_PATH="..."

python main.py \
    --iterations 10000000 \
    --dataset $DATASET \
    --max_instructions $max_instructions \
    --model $MODEL \
    --port 8000 \
    --tools_path $TOOLS_PATH \
    --max_concurrent_tasks 30 \
    --n_functions_to_sample $n_functions_to_sample \
    --n_subset_to_choose "${n_subset_array[@]}" \
    --n_in_context_examples 3 \
    --save_path $SAVE_PATH \
    --use_polisher \
    --seed_data $seed_data