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.

💁‍♂️💁💁‍♀️ Join Us on Discord!

Read this in 中文.

• [2024/3/17] Welcome to StableToolBench: A stable and reliable local toolbench server based on API response simulation. Dive deeper into the tech behind StableToolBench with paper here and explore more on the project homepage. Codes are available here.

• [2023/9/29] A new version ToolEval which is more stable and covers more models including GPT4! Please refer to ToolEval for more details. Besides, ToolLLaMA-2-7b-v2 is released with stronger tool-use capabilities. Please use the ToolLLaMA-2-7b-v2 model to reproduce our latest experimental results with the new version ToolEval.

• [2023/8/30] Data updation, with more than 120,000 solution path annotations and intact reasoning thoughts! Please find data.zip on Google Drive.

• [2023/8/8] No more hallucination! ToolLLaMA-2-7b-v1 (fine-tuned from LLaMA-2-7b) is released with lower API hallucination than ChatGPT.

• [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.

👐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 Apache License 2.0. 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:

Please download our dataset using the following link: Google Drive or Tsinghua Cloud. Notice that data_0801 is the old version data. The file structure is as follows:

├── /data/
│  ├── /instruction/
│  ├── /answer/
│  ├── /toolenv/
│  ├── /retrieval/
│  ├── /test_instruction/
│  ├── /test_query_ids/
│  ├── /retrieval_test_query_ids/
│  ├── toolllama_G123_dfs_train.json
│  └── toolllama_G123_dfs_eval.json
├── /reproduction_data/
│  ├── /chatgpt_cot/
│  ├── /chatgpt_dfs/
│  ├── ...
│  └── /toolllama_dfs/

Here are some descriptions for the data directory:

• instruction and answer: The instruction data and solution path annotation data. G1,G2, G3 refers to single-tool, intra-category multi-tool and intra-collection multi-tool data respectively. We also have an Atlas Explorer for visualization.
• toolenv: The tool environment related data, containing API jsons, API codes and API example responses.
• retrieval: The data used for tool retrieval is included in this directory.
• test_instruction and test_query_ids: We sample 200 instances from every test set. The test_instruction directory contains test queries for each test set, and the test_query_ids contains query ids of the test instances in each test set.
• retrieval_test_query_ids: This directory contains query ids of the test instances for retriever.
• toolllama_G123_dfs_train.json and toolllama_G123_dfs_eval.json: Preprocessed data that can be used to train toolllama directly and reproduce our results. For preprocessing details, 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.

Please make sure you have downloaded the necessary data and put the directory (e.g. data/) under ToolBench/, so that the following bash scripts can navigate to the related data.

We release the ToolLLaMA-2-7b-v2 which is trained on the latest version data, and ToolLLaMA-7b-v1, ToolLLaMA-7b-LoRA-v1 which are trained on the 0801 version data. All models are trained on the released dataset in a multi-task fashion. We also release the tool retriever trained under our experimental setting.

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://drive.google.com/uc?export=download&id=1XFjDxVZdUY7TXYF2yvzx3pJlS2fy78jk&confirm=yes' -O data.zip
unzip data.zip

https://drive.google.com/file/d/1XFjDxVZdUY7TXYF2yvzx3pJlS2fy78jk/view?usp=drive_link

• Data preprocessing:

export PYTHONPATH=./
python preprocess/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

• Data preprocessing, for G1_answer as an example:

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

• Our training code is based on FastChat. You can use the following command to train ToolLLaMA-7b with 2 x A100 (80GB), with our preprocessed data data/toolllama_G123_dfs_train.json. For preprocessing details, 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:

export PYTHONPATH=./
torchrun --nproc_per_node=2 --master_port=20001 toolbench/train/train_mem.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 \
    --source_model_max_length 2048 \
    --model_max_length 8192 \
    --gradient_checkpointing True \
    --lazy_preprocess True \
    --report_to none

To train lora version:

export PYTHONPATH=./
deepspeed --master_port=20001 toolbench/train/train_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 \
    --source_model_max_length 2048 \
    --model_max_length 8192 \
    --gradient_checkpointing True \
    --lazy_preprocess True \
    --deepspeed ds_configs/stage2.json \
    --report_to none

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"

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/test_instruction/G1_instruction.json \
    --output_answer_file toolllama_dfs_inference_result \
    --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/test_instruction/G1_instruction.json \
    --output_answer_file toolllama_lora_dfs_inference_result \
    --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/test_instruction/G1_instruction.json \
    --output_answer_file toolllama_lora_dfs_open_domain_inference_result \
    --toolbench_key $TOOLBENCH_KEY

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/test_instruction/G1_instruction.json \
    --output_answer_file chatgpt_dfs_inference_result \
    --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/test_instruction/G1_instruction.json \
    --output_answer_file davinci_dfs_inference_result \
    --toolbench_key $TOOLBENCH_KEY

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/test_instruction/G1_instruction.json \
    --output_answer_file chatgpt_dfs_inference_result \
    --rapidapi_key $RAPIDAPI_KEY \
    --use_rapidapi_key

To do inference with customized API(s), you should prepare the API documentation and code, then modify your query. For example, to add an API hello_world which returns a "hello world" string:

• API documentation: First generate the API documentation hello_world.json, which should follow this format:

{
    "tool_description": "Return hello world.",
    "tool_name": "hello world",
    "title": "hello world",
    "api_list": [
        {
            "name": "get_hello_world",
            "url": "",
            "description": "To get 'hello world'.",
            "method": "GET",
            "required_parameters": [],
            "optional_parameters": []
        }
    ],
    "standardized_name": "hello_world"
}

Then put it under a specific category in data/toolenv/tools/, either one of the 49 existing categories or a new category, e.g. Customized.

• API code: Create a directory naming the hello_world under Customized directory. Then write a code api.py to realize the function of the API and put it under Customized/hello_world/. The API code can be written in this format:

def get_hello_world():
    """
    To get hello world 
    """
    observation = "hello world"
    return observation

Now the file structure under data/toolenv/ should be:

├── /tools/
│  ├── /Sports/
│  │  ├── basketball.json
│  │  ├── /basketball/
│  │  │  └── api.py
│  │  └── ...
│  ├── ...
│  ├── /Customized/
│  │  ├── hello_world.json
│  │  ├── /hello_world/
│  │  │  └── api.py
└── response_examples

• Modify your query file, and the query file should follow the following format:

[
    {
        "query": "I want to get a 'hello world' string.",
        "query_id": 200001,
        "api_list": [
            {
                "category_name": "Customized",
                "tool_name": "hello world",
                "api_name": "get_hello_world"
            }
        ]
    }
]

• Finally we are free to inference with the hello_world API by running 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 /path/to/your/query/file \
    --output_answer_file /path/to/your/output/file \
    --api_customization

Currently we only support customized API usage under close-domain setting. We plan to support open-domain soon.

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.

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/

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/test_instruction/G1_instruction.json \
    --output_answer_file toolllama_lora_dfs_open_domain_result \
    --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"
}

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). More details can be found in our paper.

To validate the reliability of ChatGPT evaluator in both pass rate and win rate, we sample among four different methods (ChatGPT+ReACT, ChatGPT+DFSDT, ToolLLaMA+DFSDT and GPT4+DFSDT) to obtain solution pairs for 300 test instructions for each method. Then we engage humans to annotate the pass rate for ChatGPT+DFSDT, ToolLLaMA+DFSDT and GPT4+DFSDT, and the win rate among ChatGPT+ReACT and ChatGPT+DFSDT. Our ChatGPT evaluator demonstrates a high agreement of 87.1% in pass rate and 80.3% in win rate with human annotators. This result shows that our evaluator generates highly similar evaluation results to humans and can be viewed as a credible evaluator who simulates human evaluation on pass rate and win rate.

More details about ToolEval can be found in our paper.

Install Package (python>=3.9)

pip install -r requirements.txt

If you want to reproduce the official results, download the reproduction data reproduction_data.zip through Google Drive, unzip it and put the reproduction_data under ToolBench/data/, and skip the data preparation process.

• Data preparation. To evaluate your own model and method using ToolEval, first you need to prepare all the model predictions for the six test subsets. Create a directory naming with your model and method, e.g. chatgpt_cot then put each test set's predictions under the directory. The file sturcture of the directory should be:

├── /chatgpt_cot/
│  ├── /G1_instruction/
│  │  ├── /10160_CoT@1.json
│  │  └── ...
│  ├── /G1_tool/
│  │  ├── /10221_CoT@1.json
│  │  └── ...
│  ├── ...
│  ├── /G3_instruction/
│  │  ├── /10221_CoT@1.json
│  │  └── ...

Then preprocess the predictions by running the following commands:

export RAW_ANSWER_PATH=../../data/reproduction_data/model_predictions/
export CONVERTED_ANSWER_PATH=../../data/reproduction_data/model_predictions_converted/
export MODEL_NAME=chatgpt_cot
export METHOD=CoT
mkdir ${CONVERTED_ANSWER_PATH}/${MODEL_NAME}
for test_set in G1_instruction G1_category G1_tool G2_category G2_instruction G3_instruction
do
    answer_dir=${RAW_ANSWER_PATH}/${MODEL_NAME}/${test_set}
    output_file=${CONVERTED_ANSWER_PATH}/${MODEL_NAME}/${test_set}.json
    python convert_to_answer_format.py\
        --answer_dir ${answer_dir} \
        --method ${METHOD} \
        --output ${output_file}
done

After that, check if there are preprocessed json files for the test sets under ${CONVERTED_ANSWER_PATH}/${MODEL_NAME}. If so, you're ready to run the following evaluate process. If not, check if there is anything wrong with the model's predictions.

• OpenAI Key. Prepare your openai key to use our evaluator. The key(s) should be stored in a json file, e.g. path/to/your/openai_key_json_file.json:

[
    {
        "username": "your_user_name",
        "passwd": "your_password",
        "api_key": "your_openai_key",
        "organization": "your_organization"
    },
    ...
]

• Pass rate:

export CONVERTED_ANSWER_PATH=../../data/reproduction_data/model_predictions_converted/
export SAVE_PATH=pass_rate_results
export CANDIDATE_MODEL=chatgpt_cot
export API_POOL_FILE=path/to/your/openai_key_json_file.json

python eval_pass_rate.py \
    --converted_answer_path ${CONVERTED_ANSWER_PATH} \
    --save_path ${SAVE_PATH} \
    --reference_model ${CANDIDATE_MODEL} \
    --test_ids ../../data/test_ids/ \
    --max_eval_threads 20 \
    --evaluate_times 7

The result files will be stored under the ${SAVE_PATH}.

• Win rate. The below example take ChatGPT-ReACT as reference model and GPT4-ReACT as candidate model. Notice that you need to get both model's pass rate results first, then run the following commands to evaluate the preference result of GPT4-ReACT:

export CONVERTED_ANSWER_PATH=../../data/reproduction_data/model_predictions_converted/
export SAVE_PATH=preference_results
export PASS_TARE_PATH=pass_rate_results
export REFERENCE_MODEL=chatgpt_cot
export CANDIDATE_MODEL=gpt-4-0613_cot
export API_POOL_FILE=path/to/your/openai_key_json_file.json

python eval_preference.py \
    --converted_answer_path ${CONVERTED_ANSWER_PATH} \
    --reference_model ${REFERENCE_MODEL} \
    --output_model ${CANDIDATE_MODEL} \
    --test_ids ../../data/test_ids/ \
    --save_path ${SAVE_PATH} \
    --pass_rate_result_path ${PASS_TARE_PATH} \
    --max_eval_threads 20 \
    --use_pass_rate true \
    --evaluate_times 7

The result files will be stored under the ${SAVE_PATH}.

Please refer to ToolEval for more details.

In our main experiments, ToolLLaMA(v2) demonstrates a compelling capability to handle both single-tool and complex multi-tool instructions, which on a par with ChatGPT. Below are the main results. Win rate for each model is compared with ChatGPT-ReACT.

Pass Rate:

Win Rate: (Reference model: ChatGPT-ReACT)

• ToolLLaMA will reach GPT-4's tool-use capability.

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]

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}
}

@misc{guo2024stabletoolbench,
      title={StableToolBench: Towards Stable Large-Scale Benchmarking on Tool Learning of Large Language Models},
      author={Guo, Zhicheng and Cheng, Sijie and Wang, Hao and Liang, Shihao and Qin, Yujia and Li, Peng and Liu, Zhiyuan and Sun, Maosong and Liu, Yang},
      year={2024},
      eprint={2403.07714},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}