README.md

---
license: cc-by-nc-sa-4.0
language:
- zh
- en
size_categories:
- 100K<n<200K
task_categories:
- text-generation
---

KAgentInstruct is the instruction-tuning dataset proposed in KwaiAgents ([Github](https://github.com/KwaiKEG/KwaiAgents)), which is a series of Agent-related works open-sourced by the [KwaiKEG](https://github.com/KwaiKEG) from [Kuaishou Technology](https://www.kuaishou.com/en). It contains over 200k  agent-related instructions finetuning data (partially human-edited). Note that the dataset does not contain General-type data mentioned in the [paper](https://arxiv.org/pdf/2312.04889.pdf).


## Overall statistics of KAgentInstruct
We incorporate open-source templates ([ReACT](https://github.com/ysymyth/ReAct), [AutoGPT](https://github.com/Significant-Gravitas/AutoGPT), [ToolLLaMA](https://github.com/OpenBMB/ToolBench), [ModelScope](https://github.com/modelscope/modelscope-agent)), the KAgentSys template, and our Meta-Agent generated templates alongside the sampled queries into the experimental agent loop. This resulted in a collection of prompt-response pairs, comprising 224,137 instances, 120,917 queries and 18,005 templates, summarized in the table below.

| #Instances | #Queries | #Templates | Avg. #Steps |
|:---------:|:--------:|:----------:|:-----------:|
|  224,137   | 120,917  | 18,005    | 1.85        |

---
## Data Format


Each entry in the dataset is a dictionary with the following structure:

- `id`: A unique identifier for the entry.
- `query`: The query string.
- `source`: The origin of the data, which is one of the following: 'kwai-agent', 'meta-agent', 'autogpt', 'modelscope', 'react', 'toolllama', 'profile'.
- `functions`: A list of strings, where each string is a JSON object in string form that can be parsed into a dictionary, unless the source is 'meta-agent'.
- `function_names`: A list of function names as strings, corresponding to the functions in the `functions` list.
- `llm_prompt_response`: A list of dict, each containing:
  - `instruction`: Instruction text string.
  - `input`: Input text string.
  - `output`: Output text string.
  - `llm_name`: The name of the LLM used, either 'gpt4' or 'gpt3.5'.
  - `human_edited`: A Boolean value indicating whether the response was edited by a human.
- `extra_infos`: A dictionary containing additional useful information.

This format is designed for clarity and streamlined access to data points within the dataset.


The overall data format is as follows,
```json
{
  "id": "",
  "query": "",
  "source": "",
  "functions": [],
  "function_names": [],
  "llm_prompt_response": [
    {
        'instruction': "",
        'input': "",
        'output': "",
        'llm_name': "",
        'human_edited': bool
    },
    ...
  ],
  "extra_infos": {}
}
```
---
## How to download KAgentInstruct


You can download the KAgentInstruct through [kwaikeg/KAgentBench](https://huggingface.co/datasets/kwaikeg/KAgentInstruct/tree/main)

---

## Citation
```
@article{pan2023kwaiagents,
  author    = {Haojie Pan and
               Zepeng Zhai and
               Hao Yuan and
               Yaojia Lv and
               Ruiji Fu and
               Ming Liu and
               Zhongyuan Wang and
               Bing Qin
               },
  title     = {KwaiAgents: Generalized Information-seeking Agent System with Large Language Models},
  journal   = {CoRR},
  volume    = {abs/2312.04889},
  year      = {2023}
}
```