Vietnam History Chat Datasets (VI & EN)

Synthetic, instruction-style chat datasets about Vietnamese history (905–2025) for training and evaluating reasoning-capable chat models (GPT-OSS/Harmony, TRL, etc.).
Each sample follows a ShareGPT/ChatML-like messages schema with optional assistant reasoning in an analysis channel and a concise reply in final.

If this project helps you, please ⭐ star the repo and ❤️ Like the dataset on Hugging Face!

---

⚠️ Disclaimer

Dataset collection by AI from public source so it can make mistakes.
This dataset was collected/generated by AI from public sources and may contain errors or omissions. Please verify and curate before production use.

Tiếng Việt: Bộ dữ liệu được thu thập/tạo bởi AI từ nguồn công khai nên có thể có sai sót. Hãy kiểm chứng trước khi dùng cho mục đích sản xuất.

---

📚 Available datasets (Hugging Face)

• 1M (EN): https://huggingface.co/datasets/minhxthanh/Vietnam-History-1M-En
• 1M (VI): https://huggingface.co/datasets/minhxthanh/Vietnam-History-1M-Vi
• 500K (VI): https://huggingface.co/datasets/minhxthanh/Vietnam-History-500K-Vi
• 500K (EN): https://huggingface.co/datasets/minhxthanh/Vietnam-History-500K-En
• 200K (VI): https://huggingface.co/datasets/minhxthanh/Vietnam-History-200K-Vi
• 200K (EN): https://huggingface.co/datasets/minhxthanh/Vietnam-History-200K-EN
• 100K (VI): https://huggingface.co/datasets/minhxthanh/Vietnam-History-100K_Vi
• 100K (EN): https://huggingface.co/datasets/minhxthanh/VietNam-History-100K_EN
• 50K (VI): https://huggingface.co/datasets/minhxthanh/Vietnam-History-50K
• 25K (VI): https://huggingface.co/datasets/minhxthanh/VietNam-History-25K
• 15K (VI): https://huggingface.co/datasets/minhxthanh/Vietnam-History-15k

See the author profile for activity and updates: https://huggingface.co/minhxthanh

---

🧠 Content coverage

From Khúc/Ngô (905) → Đinh – Tiền Lê – Lý – Trần – Hồ – Minh thuộc – Lê sơ – Mạc – Lê trung hưng (Trịnh–Nguyễn) – Tây Sơn – Nguyễn – thời Pháp thuộc – 1945–1975 – Đổi Mới – hội nhập – đương đại (đến 2025).
Prompt styles include year-based, event summaries, person-focused, why/how causality, and context → developments → outcome.

---

🗂️ Data schema

Each row is a JSON object with a single messages list:

{
  "messages": [
    {"role": "system", "content": "You are a Vietnam history assistant. Answer concisely."},
    {"role": "user", "content": "What happened in 938?"},
    {"role": "assistant", "channel": "analysis", "content": "Identify the key event..."},
    {"role": "assistant", "channel": "final", "content": "In 938, the Battle of Bạch Đằng... ended Chinese rule."}
  ]
}

• role ∈ {system, user, assistant}.

• content: string.

• Optional channel for assistant:

  • analysis → chain-of-thought / reasoning (present in most samples).
  • final → short answer shown to the user.

• A minority of records are final-only (no analysis) to balance concise vs. reasoning behavior.

VI example

{
  "messages": [
    {"role":"system","content":"Bạn là trợ lý lịch sử Việt Nam."},
    {"role":"user","content":"Tóm tắt Trận Bạch Đằng (1288)."},
    {"role":"assistant","channel":"analysis","content":"Xác định bối cảnh kháng Nguyên lần 3..."},
    {"role":"assistant","channel":"final","content":"Năm 1288, Trần Hưng Đạo nhử địch vào bãi cọc trên sông Bạch Đằng và tiêu diệt thủy quân Nguyên."}
  ]
}

---

🚀 Quick start (🤗 Datasets)

from datasets import load_dataset

ds = load_dataset("minhxthanh/Vietnam-History-500K-Vi", split="train")
print(ds[0]["messages"][0]["role"], ":", ds[0]["messages"][0]["content"])

Tip: Large splits (e.g., 500K) may stream/convert to Parquet on first load.

---

🦥 Harmony / GPT-OSS (Unsloth) workflow

These datasets map cleanly to Harmony chat format for GPT-OSS training:

# pip install unsloth
from unsloth_zoo import encode_conversations_with_harmony as enc
from datasets import load_dataset

ds = load_dataset("minhxthanh/Vietnam-History-200K-EN", split="train")

def to_harmony(example):
    return {"input_ids": enc(example["messages"])["input_ids"]}

ds_tok = ds.map(to_harmony, remove_columns=ds.column_names)

Use ds_tok directly in your SFT trainer.

---

📦 Shards & sizes

• Each HF dataset provides a single train split with the indicated size (e.g., 500k, 200k, 100k, …).
• The 500K releases may be internally sharded for easier hosting; loading through 🤗 Datasets handles this automatically.

---

🔐 License

Each dataset page declares an MIT license (check the HF sidebar for the authoritative license for that dataset). Please attribute the dataset URLs when publishing models trained on them.

---

🤝 Contributing

• Found an error or want to add more prompts? Open an issue or PR.
• If this project is useful, please give it a ⭐ on GitHub and ❤️ Like on the HF dataset page.

---

✍️ Author

• Hugging Face: https://huggingface.co/minhxthanh
• Model example: minhxthanh/DeepSeek-R1-Distill-Llama-8B-VN-History (VN-history-focused 8B model)