An open platform for training, serving, and evaluating large language model for tool learning.

Model • Data Release • Web Demo • Tool Eval • Paper • Citation

🔨This project (ToolLLM) aims to construct open-source, large-scale, high-quality instruction tuning SFT data to facilitate the construction of powerful LLMs with general tool-use capability. We aim to empower open-source LLMs to master thousands of diverse real-world APIs. We achieve this by collecting a high-quality instruction-tuning dataset. It is constructed automatically using the latest ChatGPT (gpt-3.5-turbo-16k), which is upgraded with enhanced function call capabilities. We provide the dataset, the corresponding training and evaluation scripts, and a capable model ToolLLaMA fine-tuned on ToolBench.

💁‍♂️💁💁‍♀️Joint Us on Discord!

What's New

• [2023/8/4] We provide RapidAPI backend service to free you from using your own RapidAPI key and subscribing the APIs. Please fill out our form. We will review it as soon as possible and send you the ToolBench key to get start on it!

• [2023/8/1] Our paper is released.

• [2023/7/27] New version ToolBench is released.

✨Here is an overview of the dataset construction, training, and evaluation.

✨✨Features:

• API Collection: we gather 16464 representational state transfer (REST) APIs from RapidAPI, a platform that hosts massive real-world APIs provided by developers.
• Instruction Generation: we curate instructions that involve both single-tool and multi-tool scenarios.
• Answer Annotation: we develop a novel depth-first search based decision tree (DFSDT) to bolster the planning and reasoning ability of LLMs, which significantly improves the annotation efficiency and successfully annotates those complex instructions that cannot be answered with CoT or ReACT. We provide responses that not only include the final answer but also incorporate the model's reasoning process, tool execution, and tool execution results.
• API Retriver: we incorporate API retrieval to equip ToolLLaMA with open-domain tool-using abilities.
• All the data is automatically generated by OpenAI API and filtered by us, the whole data creation process is easy to scale up.

We also provide A demo of using ToolLLaMA

Currently, our ToolLLaMA has reached the performance of ChatGPT (turbo-16k) in tool use, in the future, we will continually improve the data quality and increase the coverage of real-world tools.

Here is the Old version of ToolBench.

Data

👐ToolBench is intended solely for research and educational purposes and should not be construed as reflecting the opinions or views of the creators, owners, or contributors of this dataset. It is distributed under CC BY NC 4.0 License. Below is the statistics of the data :

We crawl 16000+ real-world APIs from RapidAPI, and curate realistic human instructions that involve them. Below we present a hierarchy of RapidAPI and our instruction generation process.

ToolBench contains both single-tool and multi-tool scenarios. The multi-tool scenarios can be further categorized into intra-category multi-tool and intra-collection multi-tool. We utilize DFSDT method for all scenarios to our data creation. Here is an illustration for the data creation process using DFSDT method:

Data Release

Please download our dataset using the following link: Google Drive or Tsinghua Cloud.

• G1,G2, G3data refers to single-tool, intra-category multi-tool and intra-collection multi-tool data respectively. We also have an Atlas Explorer for visualization.
• We split the G1, G2 and G3 data into train, eval and test parts respectively and combine the train data for training in our main experiments. toolllama_G123_dfs_train.json refers to the combined train data.
• The tool environment related data is in toolenv directory.
• We sample 100 instances from every test set. The test_query_ids directory contains query ids of the test instances in each test set.
• The data used for tool retrieval is included in the retrieval directory.

🤖Model

We release the ToolLLaMA-7b and ToolLLaMA-7b-LoRA models, which are both trained on the released dataset in a multi-task fashion. We also release the tool retriever trained under our experimental setting.

🚀Fine-tuning

Install

Clone this repository and navigate to the ToolBench folder.

git clone git@github.com:OpenBMB/ToolBench.git
cd ToolBench

Install Package (python>=3.9)

pip install -r requirements.txt

or for ToolEval only

pip install -r toolbench/tooleval/requirements.txt

Prepare the data and tool environment:

wget --no-check-certificate 'https://docs.google.com/uc?export=download&id=1Vis-RxBstXLKC1W1agIQUJNuumPJrrw0&confirm=yes' -O data.zip
unzip data.zip

Training Retriever

• Data preprocessing:

export PYTHONPATH=./
python data/preprocess_retriever_data.py \
    --query_file data/instruction/G1_query.json \
    --index_file data/test_query_ids/G1_instruction_test_query_ids.json \
    --dataset_name G1 \
    --output_dir data/retrieval/G1

• Then run the following command to train the tool retriever:

export PYTHONPATH=./
python toolbench/retrieval/train.py \
    --data_path data/retrieval/G1/ \
    --model_name bert-base-uncased \
    --output_path retrieval_model \
    --num_epochs 5 \
    --train_batch_size 32 \
    --learning_rate 2e-5 \
    --warmup_steps 500 \
    --max_seq_length 256

Training ToolLLaMA

Our training code is based on FastChat. You can use the following command to train ToolLLaMA-7b with 2 x A100 (80GB), with the preprocessed data in our data link:

export PYTHONPATH=./
torchrun --nproc_per_node=2 --master_port=20001 toolbench/train/train_long_seq.py \
    --model_name_or_path huggyllama/llama-7b  \
    --data_path  data/toolllama_G123_dfs_train.json \
    --eval_data_path  data/toolllama_G123_dfs_eval.json \
    --conv_template tool-llama-single-round \
    --bf16 True \
    --output_dir toolllama \
    --num_train_epochs 2 \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 2 \
    --gradient_accumulation_steps 8 \
    --evaluation_strategy "epoch" \
    --prediction_loss_only \
    --save_strategy "epoch" \
    --save_total_limit 8 \
    --learning_rate 5e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.04 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --fsdp "full_shard auto_wrap" \
    --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
    --tf32 True \
    --model_max_length 8192 \
    --gradient_checkpointing True \
    --lazy_preprocess True \
    --report_to none

You can also preprocess and split the data in your own way with this command:

export PYTHONPATH=./
python preprocess/preprocess_toolllama_data.py \
    --tool_data_dir data/answer/G1_answer \
    --method DFS_woFilter_w2 \
    --output_file data/answer/toolllama_G1_dfs.json

To train lora version:

export PYTHONPATH=./
deepspeed --master_port=20001 toolbench/train/train_long_seq_lora.py \
    --model_name_or_path huggyllama/llama-7b  \
    --data_path  data/toolllama_G123_dfs_train.json \
    --eval_data_path  data/toolllama_G123_dfs_eval.json \
    --conv_template tool-llama-single-round \
    --bf16 True \
    --output_dir toolllama_lora \
    --num_train_epochs 5 \
    --per_device_train_batch_size 4 \
    --per_device_eval_batch_size 2 \
    --gradient_accumulation_steps 2 \
    --evaluation_strategy "epoch" \
    --prediction_loss_only \
    --save_strategy "epoch" \
    --save_total_limit 8 \
    --learning_rate 5e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.04 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --model_max_length 8192 \
    --gradient_checkpointing True \
    --lazy_preprocess True \    
    --deepspeed ds_configs/stage2.json \
    --report_to none

Inference With Our RapidAPI Server

Please fill out the form first and after reviewing we will send you the toolbench key. Then prepare your toolbench key by:

export TOOLBENCH_KEY="your_toolbench_key"

For ToolLLaMA

To inference with ToolLLaMA, run the following commands:

export PYTHONPATH=./
python toolbench/inference/qa_pipeline.py \
    --tool_root_dir data/toolenv/tools/ \
    --backbone_model toolllama \
    --model_path ToolBench/ToolLLaMA-7b \
    --max_observation_length 1024 \
    --observ_compress_method truncate \
    --method DFS_woFilter_w2 \
    --input_query_file data/instruction/inference_query_demo.json \
    --output_answer_file data/answer/toolllama_dfs \
    --toolbench_key $TOOLBENCH_KEY

For ToolLLaMA-LoRA:

export PYTHONPATH=./
python toolbench/inference/qa_pipeline.py \
    --tool_root_dir data/toolenv/tools/ \
    --backbone_model toolllama \
    --model_path huggyllama/llama-7b \
    --lora \
    --lora_path /path/to/your/downloaded/ToolLLaMA-7b-LoRA \
    --max_observation_length 1024 \
    --observ_compress_method truncate \
    --method DFS_woFilter_w2 \
    --input_query_file data/instruction/inference_query_demo.json \
    --output_answer_file data/answer/toolllama_lora_dfs \
    --toolbench_key $TOOLBENCH_KEY

For ToolLLaMA-LoRA under open-domain setting, run:

export PYTHONPATH=./
python toolbench/inference/qa_pipeline_open_domain.py \
    --tool_root_dir data/toolenv/tools/ \
    --corpus_tsv_path data/retrieval/G1/corpus.tsv \
    --retrieval_model_path /path/to/your/retrival_model \
    --retrieved_api_nums 5 \
    --backbone_model toolllama \
    --model_path huggyllama/llama-7b \
    --lora \
    --lora_path /path/to/your/toolllama_lora \
    --max_observation_length 1024 \
    --observ_compress_method truncate \
    --method DFS_woFilter_w2 \
    --input_query_file data/instruction/inference_query_demo_open_domain.json \
    --output_answer_file data/answer/toolllama_lora_dfs_open_domain \
    --toolbench_key $TOOLBENCH_KEY

For OpenAI Models

To use ChatGPT, run:

export TOOLBENCH_KEY=""
export OPENAI_KEY=""
export PYTHONPATH=./
python toolbench/inference/qa_pipeline.py \
    --tool_root_dir data/toolenv/tools/ \
    --backbone_model chatgpt_function \
    --openai_key $OPENAI_KEY \
    --max_observation_length 1024 \
    --method DFS_woFilter_w2 \
    --input_query_file data/instruction/inference_query_demo.json \
    --output_answer_file data/answer/chatgpt_dfs \
    --toolbench_key $TOOLBENCH_KEY

To use Text-Davinci-003, run:

export TOOLBENCH_KEY=""
export OPENAI_KEY=""
export PYTHONPATH=./
python toolbench/inference/qa_pipeline.py \
    --tool_root_dir data/toolenv/tools/ \
    --backbone_model davinci \
    --openai_key $OPENAI_KEY \
    --max_observation_length 1024 \
    --method DFS_woFilter_w2 \
    --input_query_file data/instruction/inference_query_demo.json \
    --output_answer_file data/answer/davinci_dfs \
    --toolbench_key $TOOLBENCH_KEY

Inference With Your Own RapidAPI Account

To do inference with customized RapidAPI account, pass your rapidapi key through rapidapi_key and specify the use_rapidapi_key argument in the script:

export RAPIDAPI_KEY=""
export OPENAI_KEY=""
export PYTHONPATH=./
python toolbench/inference/qa_pipeline.py \
    --tool_root_dir data/toolenv/tools/ \
    --backbone_model chatgpt_function \
    --openai_key $OPENAI_KEY \
    --max_observation_length 1024 \
    --method DFS_woFilter_w2 \
    --input_query_file data/instruction/inference_query_demo.json \
    --output_answer_file data/answer/chatgpt_dfs \
    --rapidapi_key $RAPIDAPI_KEY \
    --use_rapidapi_key

Setting up and running the interface

ToolBench contains a Web UI based on Chatbot UI, forked to include the use of tools in the interface. It comes in two parts: the backend server, and chatbot-ui-toolllama. Here is a video demo.

Web UI

git clone https://github.com/lilbillybiscuit/chatbot-ui-toolllama
cd chatbot-ui-toolllama
npm install
npm run dev

The app will be available on http://localhost:3000/

Backend server

export PYTHONPATH=./
python toolbench/inference/toolbench_server.py \
    --tool_root_dir data/toolenv/tools/ \
    --corpus_tsv_path data/retrieval/G1/corpus.tsv \
    --retrieval_model_path /path/to/your/retrival_model \
    --retrieved_api_nums 5 \
    --backbone_model toolllama \
    --model_path huggyllama/llama-7b \
    --lora \
    --lora_path /path/to/your/toolllama_lora \
    --max_observation_length 1024 \
    --method DFS_woFilter_w2 \
    --input_query_file data/instruction/inference_query_demo_open_domain.json \
    --output_answer_file data/answer/toolllama_lora_dfs_open_domain \
    --rapidapi_key $RAPIDAPIKEY

This server will be available on http://localhost:5000/. To start a request, call http://localhost:5000/stream with a GET or POST request containing a JSON object with the following fields:

{
    "text": "What is the weather in New York today?",
    "top_k": 5,
    "method": "DFS_woFilter_w2"
}

ToolEval

By fine-tuning LLaMA on ToolBench, we obtain ToolLLaMA. Considering that human evaluation can be time-consuming, we follow AlpacaEval to develop an efficient machine evaluator ToolEval, which incorporates two evaluation metrics:

• Pass Rate: Calculates the proportion of successfully completing an instruction within limited OpenAI API calls.
• Preference: Measured by comparing two answers (action sequences) for a given instruction. We pre-define a set of criteria for a better answer, which are organized as prompts for ChatGPT. We provide the test instruction and two candidate answers to the evaluator and obtain its preference. We evaluate each answer pair multiple times to improve the reliability of our system. Then we calculate the Win Rate (percentage of being preferred by the evaluator) and Standard Error (the standard error of the Win Rate). More details can be found in our paper.

To validate the effectiveness of the metric Preference, we sample among three different methods (ChatGPT+ReACT, GPT4+ReACT, and ChatGPT+DFSDT) to obtain answer pairs for 600 test instructions. Then we engage humans to annotate human preference for them (4 annotations for each answer pair, 2400 annotations in total). Our automatic evaluator, developed using ChatGPT, demonstrates a significant correlation of 75.8% with human annotators. We also obtain the agreement among different human annotators 83.54%, and the agreement between humans and our evaluator 80.21%.

More details about ToolEval can be found in our paper.

Evaluation with ToolEval

To evaluate a model on G1-Inst. test set, for example, run the following commands.

• Pass rate:

python toolbench/tooleval/pass_rate.py --answer_dir data/answer/toolllama_dfs/G1_instruction

• Win rate (Reference model: ChatGPT-ReACT):

export OPENAI_KEY=""
export REF_MODEL_DATA="data/answer/chatgpt_cot/G1_instruction"
export REF_MODEL_METHOD="CoT"
export TEST_MODEL_DATA="data/answer/toolllama_dfs/G1_instruction"
export TEST_MODEL_METHOD="DFS"
python ./toolbench/tooleval/convert_to_answer_format.py \
    --method CoT \
    --answer_dir $REF_MODEL_DATA \
    --output ${REF_MODEL_DATA}_converted

python ./toolbench/tooleval/convert_to_answer_format.py \
    --method DFS \
    --answer_dir $TEST_MODEL_DATA \
    --output ${TEST_MODEL_DATA}_converted

python ./toolbench/tooleval/automatic_eval_sample.py \
    --output ${REF_MODEL_DATA}_converted \
    --ref_output ${TEST_MODEL_DATA}_converted \
    --method $REF_MODEL_METHOD \
    --use_existed_output

📊 Model Experiments Results

In our main experiments, ToolLLaMA demonstrates a compelling capability to handle both single-tool and complex multi-tool instructions. Below are the main results compared with ChatGPT and Text-Davinci-003.

Pass Rate:

Win Rate: (Reference model: ChatGPT-DFSDT)

TODO

• ToolLLaMA will reach GPT-4's tool-use capability.
• We will train a ToolLLaMa-2

Resources of Tool Learning

With the powerful capabilities of foundation models, we are eager to see their applications in manipulating various tools. For more resources, please refer to the following:

• BMTools. [Project]

• Tool Learning Survey. [Paper]

• Tool Learning Paper List. [Project]

• WebCPM. [Paper]

Citation

Feel free to cite us if you like ToolBench.

@misc{qin2023toolllm,
      title={ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs}, 
      author={Yujia Qin and Shihao Liang and Yining Ye and Kunlun Zhu and Lan Yan and Yaxi Lu and Yankai Lin and Xin Cong and Xiangru Tang and Bill Qian and Sihan Zhao and Runchu Tian and Ruobing Xie and Jie Zhou and Mark Gerstein and Dahai Li and Zhiyuan Liu and Maosong Sun},
      year={2023},
      eprint={2307.16789},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}

@misc{qin2023tool,
      title={Tool Learning with Foundation Models}, 
      author={Yujia Qin and Shengding Hu and Yankai Lin and Weize Chen and Ning Ding and Ganqu Cui and Zheni Zeng and Yufei Huang and Chaojun Xiao and Chi Han and Yi Ren Fung and Yusheng Su and Huadong Wang and Cheng Qian and Runchu Tian and Kunlun Zhu and Shihao Liang and Xingyu Shen and Bokai Xu and Zhen Zhang and Yining Ye and Bowen Li and Ziwei Tang and Jing Yi and Yuzhang Zhu and Zhenning Dai and Lan Yan and Xin Cong and Yaxi Lu and Weilin Zhao and Yuxiang Huang and Junxi Yan and Xu Han and Xian Sun and Dahai Li and Jason Phang and Cheng Yang and Tongshuang Wu and Heng Ji and Zhiyuan Liu and Maosong Sun},
      year={2023},
      eprint={2304.08354},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}