Home

Awesome

ToolAlpaca: Generalized Tool Learning for Language Models with 3000 Simulated Cases

arXiv

<div align=center><img src="./figures/ToolAlpaca.png" width="400px" /></div>

ToolAlpaca is a framework designed for learning generalized tool-use abilities in compact language models with minimal human supervision. It addresses the challenge of tool learning by generating a tool-use corpus via a multi-agent simulation environment, providing 3.9k tool-use instances from more than 400 tools.

<div align=center><img src="./figures/pipeline.png" width="500px" /></div>

Data

Dataset list:

Data format:

{
  "Name": "name, from public-apis",
  "Description": "description, from public-apis",
  "Category": "category, from public-apis",
  "Introduction": "introduction, generated by LLM",
  "Functions": "NLDocumentation in paper v1, generated by LLM",
  "Documentation": "str(json), OpenAPI Specification documentation, generated by LLM",
  "NLDocumentation": "natural language documentation, similar to Functions, converted from Documentation",
  "Function_Description": "each functions description in NLDocumentation",
  "Function_Projection": "function to HTTP request method",
  "Instructions": "instructions, generated by LLM",
  "Instances": [
    {
      "input": "use's init instruction, from use agent",
      "output": "final output, from assistant agent",
      "Final Thought": "the final thought before output, from assistant agent",
      "intermediate_steps": [
        [
          [
            "action, from assistant agent",
            "action input, str(json), from assistant agent",
            "thought + action + action input, assistant agent's output"
          ]
          "bbservation, from [user agent, type check python code, tool executor agent]"
        ]
      ]
    }
  ]
}

Dataset Generation

git clone git@github.com:tangqiaoyu/ToolAlpaca.git
cd ToolAlpaca
pip install -r requirements.txt
python tool_maker/preprocess_public_apis.py -api data/public_apis.json
export PYTHONPATH=$PYTHONPAT:$(pwd)
export OPENAI_API_KEY=""

python tool_maker/get_elements.py -api data/public_apis.json -out ./data

python tool_maker/natural_language_documentation.py -api ./data/api_data.json
python instance_generation/instruction.py -api ./data/api_data.json -out ./data

python instance_generation/simulator.py -api ./data/api_data.json

python instance_generation/generation.py -api ./data/api_data.json -out ./data --use_cache

Train

To train Toolapaca, we need to create a prompt to organize the dataset in a format that the standard SFT training code can read, similar to what is done in build_dataset.py. Afterward, we can proceed with training using the standard SFT method, only optimizing the loss on thought, action, and action input.

deepspeed --num_gpus=2 --master_port=12345 train.py \
    --deepspeed ${deepspeed config path} \
    --model_name_or_path ${path to base model like vicuna-7b}  \
    --data_path ${data path} \
    --bf16 True \
    --output_dir outputs/vicuna-7b-toolalpaca/ \
    --num_train_epochs 3 \
    --per_device_train_batch_size 32 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 2 \
    --evaluation_strategy "no" \
    --save_strategy "epoch" \
    --save_total_limit 10 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --lazy_preprocess True

You can Find our models on huggingface hub: ToolAlpaca-7B, ToolAlpaca-13B.

Evaluation

# start the api simulator
python instance_generation/simulator.py -api ./data/eval_simulated.json

# get LLM outputs
python instance_generation/generation.py \
  -api ./data/eval_simulated.json \
  -out ./eval \
  -llm TangQiaoYu/ToolAlpaca-13B \
  --agent_prompt test_v1 \
  --use_cache

# evaluation with LLM like GPT-4
python evaluation.py -api ${api_data_path} -out ./eval
python instance_generation/generation.py \
  -api ./data/eval_real.json \
  -out ./data \
  -llm TangQiaoYu/ToolAlpaca-13B \
  --agent_prompt test_v1 \
  --real

python evaluation.py -api ${api_data_path} -out ./eval

Citation

If you find our work helpful, please cite as

@misc{tang2023toolalpaca,
      title={ToolAlpaca: Generalized Tool Learning for Language Models with 3000 Simulated Cases}, 
      author={Qiaoyu Tang and Ziliang Deng and Hongyu Lin and Xianpei Han and Qiao Liang and Le Sun},
      year={2023},
      eprint={2306.05301},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}