wannaphong/typhoon-s-sovereign-capability-dataset

---
pretty_name: Typhoon-S Instruct Post-Training
tags:
- reinforcement fine-tuning
- tool-use
- thai
- english
- sovereign-ai
task_categories:
- text-generation
language:
- th
- en
license: odc-by
---

# Typhoon-S Training Assets

Training and evaluation datasets for Section 3, Thai language models used in the Typhoon-S project.

## Datasets

**NitiBench (Legal Domain)**
- `nitibench_train_rl.parquet` - RL training set (8,211 examples)
- `nitibench_train_pretrain.parquet` - Pretrain set (3,648 examples)
- `nitibench_train_sft.parquet` - SFT set (3,648 examples)
- `nitibench_test.parquet` - Test set (373 examples) (10% of https://huggingface.co/datasets/VISAI-AI/nitibench ccl split)
- `nitibench_train_rl_agent.parquet` - Agent RL training (8,211 examples)

Original source
- https://huggingface.co/datasets/airesearch/WangchanX-Legal-ThaiCCL-RAG
- https://huggingface.co/datasets/VISAI-AI/nitibench

**MIRAGE (General Domain)**
- `mirage_train_rl.parquet` - RL training set
- `mirage_train_pretrain.parquet` - Pretrain set
- `mirage_test.parquet` - Test set

Original source
- https://huggingface.co/datasets/nthakur/mirage-bench-instruct
- https://huggingface.co/datasets/nthakur/mirage-bench

## Usage

```python
from datasets import load_dataset

# Load a single file
dataset = load_dataset("typhoon-ai/typhoon-s-sovereign-capability-dataset", data_files="nitibench_train_rl.parquet")

# Load multiple files
dataset = load_dataset(
"typhoon-ai/typhoon-s-sovereign-capability-dataset",
data_files={
"train": "nitibench_train_rl.parquet",
"test": "nitibench_test.parquet"
}
)
```

## More Information

Please see for more details: https://github.com/scb-10x/typhoon-s

## Citation

If you use this dataset, please cite the dataset repository and the associated Typhoon-S technical report:

```bibtex
@misc{pipatanakul2026typhoonsminimalopenposttraining,
title={Typhoon-S: Minimal Open Post-Training for Sovereign Large Language Models},
author={Kunat Pipatanakul and Pittawat Taveekitworachai},
year={2026},
eprint={2601.18129},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2601.18129},
}
```

数据集展示名称：Typhoon-S 指令后训练（Typhoon-S Instruct Post-Training）
标签：
- 强化微调（reinforcement fine-tuning）
- 工具使用（tool-use）
- 泰语（thai）
- 英语（english）
- 主权AI（sovereign-ai）
任务类别：
- 文本生成（text-generation）
语言：
- 泰语（th）
- 英语（en）
许可协议：odc-by
---

# 台风-S 训练资源
本数据集为台风-S项目第3章节所用泰语大语言模型（Large Language Model）的训练与评测数据集。

## 数据集
### NitiBench（法律领域）
- `nitibench_train_rl.parquet`：强化学习训练集（共8211条样本）
- `nitibench_train_pretrain.parquet`：预训练集（共3648条样本）
- `nitibench_train_sft.parquet`：监督微调（Supervised Fine-Tuning, SFT）训练集（共3648条样本）
- `nitibench_test.parquet`：测试集（共373条样本），为https://huggingface.co/datasets/VISAI-AI/nitibench数据集的CCL划分子集（占比10%）
- `nitibench_train_rl_agent.parquet`：智能体强化学习训练集（共8211条样本）

原始数据来源：
- https://huggingface.co/datasets/airesearch/WangchanX-Legal-ThaiCCL-RAG
- https://huggingface.co/datasets/VISAI-AI/nitibench

### MIRAGE（通用领域）
- `mirage_train_rl.parquet`：强化学习训练集
- `mirage_train_pretrain.parquet`：预训练集
- `mirage_test.parquet`：测试集

原始数据来源：
- https://huggingface.co/datasets/nthakur/mirage-bench-instruct
- https://huggingface.co/datasets/nthakur/mirage-bench

## 使用方法
python
from datasets import load_dataset

# 加载单个数据文件
dataset = load_dataset("typhoon-ai/typhoon-s-sovereign-capability-dataset", data_files="nitibench_train_rl.parquet")

# 加载多个数据文件
dataset = load_dataset(
"typhoon-ai/typhoon-s-sovereign-capability-dataset",
data_files={
"train": "nitibench_train_rl.parquet",
"test": "nitibench_test.parquet"
}
)

## 更多信息
如需获取更多细节，请访问：https://github.com/scb-10x/typhoon-s

## 引用说明
若您使用本数据集，请引用本数据集仓库以及关联的台风-S技术报告：
bibtex
@misc{pipatanakul2026typhoonsminimalopenposttraining,
title={Typhoon-S: Minimal Open Post-Training for Sovereign Large Language Models},
author={Kunat Pipatanakul and Pittawat Taveekitworachai},
year={2026},
eprint={2601.18129},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2601.18129},
}