RikkaBotan/FineDataset_SFT
收藏Hugging Face2025-11-23 更新2025-12-20 收录
下载链接：
https://hf-mirror.com/datasets/RikkaBotan/FineDataset_SFT
下载链接
链接失效反馈官方服务：
资源简介：
---
license: cc-by-4.0
dataset_info:
features:
- name: messages
list:
- name: role
dtype: string
- name: content
dtype: string
splits:
- name: train
num_bytes: 13070934724
num_examples: 1464841
download_size: 6161254076
dataset_size: 13070934724
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---
# Posttraining Dataset
This document describes the construction of the posttraining dataset.
The dataset is built by sampling multiple subsets from **NVIDIA’s Nemotron-Post-Training-Dataset-v2**, following a carefully designed token ratio to balance Japanese text, chat-style dialogs, code, STEM, and math reasoning data.
---
## 📚 Overview
The dataset blends **Japanese-focused data**, **instruction/chat**, **coding**, **STEM**, and **mathematical reasoning**.
The dataset is constructed from the following Hugging Face dataset:
- **`nvidia/Nemotron-Post-Training-Dataset-v2`**
Each subset is sampled proportionally to the target token ratio described below.
---
## 📊 Token Ratios
The dataset uses the following token distribution:
| Subset | Tokens |
|-------------------|------------------|
| `nemotron_ja` | 900,000,000 |
| `nemotron_chat` | 600,000,000 |
| `nemotron_code` | 600,000,000 |
| `nemotron_stem` | 600,000,000 |
| `nemotron_math` | 300,000,000 |
Total tokens: **3.0B**
This ratio gives Saint Iberis strong general-purpose Japanese capability while also ensuring high competency in reasoning, coding, and multi-domain problem solving.
---
## 🧩 Dataset Configuration
The dataset subsets are derived using the following configuration:
```bash
"nemotron_ja": {
"hf_name": "nvidia/Nemotron-Post-Training-Dataset-v2",
"subset": "multilingual_ja",
}
"nemotron_chat": {
"hf_name": "nvidia/Nemotron-Post-Training-Dataset-v2",
"subset": "chat",
}
"nemotron_code": {
"hf_name": "nvidia/Nemotron-Post-Training-Dataset-v2",
"subset": "code",
}
"nemotron_stem": {
"hf_name": "nvidia/Nemotron-Post-Training-Dataset-v2",
"subset": "stem",
}
"nemotron_math": {
"hf_name": "nvidia/Nemotron-Post-Training-Dataset-v2",
"subset": "math",
}
```
# 🌸 About us
Japanese independent researcher having shy and pampered personality. Twin-tail hair is a charm point. Interested in nlp. Usually using python and C.
提供机构：
RikkaBotan
