p988744/eland-official-doc-zh

Eland Official Document Chinese Dataset

A Chinese dataset for training models to convert colloquial text into formal Taiwan government official document format.

Dataset Description

This dataset contains:

1. SFT Data: Colloquial input → Formal official document output pairs
2. ORPO Preference Data: Chosen (correct terminology) vs Rejected (incorrect terminology) pairs

Dataset Statistics

Split	Samples	Description
train	80	SFT training samples
valid	10	Validation samples
test	10	Test samples
orpo_train	498	ORPO preference pairs

Supported Document Types

Type	Chinese	Description
函	Official Letter	Inter-agency correspondence
簽	Internal Proposal	Internal reporting to superiors
公告	Public Announcement	Public notices
書函	Semi-formal Letter	Less formal communication
開會通知	Meeting Notice	Meeting notifications

Document Levels & Terminology

Level	Chinese	Addressee	Closing
上行	Upward	鈞部/鈞府	請鑒核/請核示
平行	Horizontal	貴部/貴公司	請查照/請惠復
下行	Downward	該部/該單位	希查照/希遵辦

Data Format

SFT Format (sft config)

Note: The output field is stored as a JSON string (not a nested object) to ensure consistent Parquet schema.

{
  "input": "口語化的請求內容...",
  "doc_type": "函",
  "level": "上行",
  "domain": "行政管理",
  "output": "{\"subject\": \"主旨內容\", \"explanation\": [\"說明一\", \"說明二\"], ...}"
}

After parsing output:

{
  "subject": "主旨內容",
  "explanation": ["說明一", "說明二"],
  "action": ["擬辦事項"],
  "formatted_text": "完整格式化公文"
}

ORPO Format (orpo config)

{
  "prompt": "請將以下口語轉換為上行公文：...",
  "chosen": "主旨：檢陳...敬請惠予補助。\n\n說明：依據鈞部...",
  "rejected": "主旨：檢陳...請惠予補助。\n\n說明：依據貴部..."
}

Usage

from datasets import load_dataset
import json

# Load SFT config (for supervised fine-tuning)
sft_dataset = load_dataset("p988744/eland-official-doc-zh", "sft")

# Load ORPO config (for preference learning)
orpo_dataset = load_dataset("p988744/eland-official-doc-zh", "orpo")

# Example - SFT (output is JSON string)
sample = sft_dataset["train"][0]
print(f"Input: {sample['input']}")
output = json.loads(sample['output'])  # Parse JSON string
print(f"Output: {output['formatted_text']}")

# Example - ORPO
sample = orpo_dataset["train"][0]
print(f"Prompt: {sample['prompt']}")
print(f"Chosen: {sample['chosen']}")
print(f"Rejected: {sample['rejected']}")

Token Statistics

⚠️ Note: This dataset has two configs with different schemas.

SFT Config

Split	Samples	Estimated Tokens
train	80	~58K
valid	10	~7K
test	10	~7K

ORPO Config

Split	Samples	Estimated Tokens
train	498	~180K

Token counts estimated using Qwen3-4B tokenizer.

Calculate Tokens

from datasets import load_dataset
from transformers import AutoTokenizer
import json

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")

def count_tokens(text):
    return len(tokenizer.encode(str(text), add_special_tokens=False))

# SFT config (input + output JSON string)
sft_dataset = load_dataset("p988744/eland-official-doc-zh", "sft")
sft_tokens = 0
for sample in sft_dataset["train"]:
    sft_tokens += count_tokens(sample["input"])
    # Parse JSON string and count formatted_text
    output = json.loads(sample["output"])
    sft_tokens += count_tokens(output.get("formatted_text", ""))
print(f"SFT train tokens: {sft_tokens:,}")

# ORPO config (prompt + chosen + rejected)
orpo_dataset = load_dataset("p988744/eland-official-doc-zh", "orpo")
orpo_tokens = 0
for sample in orpo_dataset["train"]:
    orpo_tokens += count_tokens(sample["prompt"])
    orpo_tokens += count_tokens(sample["chosen"])
    orpo_tokens += count_tokens(sample["rejected"])
print(f"ORPO train tokens: {orpo_tokens:,}")

Training

This dataset was used to train the official document generation model using ORPO (Odds Ratio Preference Optimization):

• Base Model: Qwen/Qwen3-4B
• Method: ORPO
• Performance: 100% terminology accuracy across all document levels

Related Models

• eland-official-doc-zh - LoRA Adapter
• eland-official-doc-zh-gguf - GGUF for Ollama
• eland-official-doc-zh-vllm - Merged for vLLM

License

Apache 2.0

Citation

@dataset{eland_official_doc_zh_data,
  title={Eland Official Document Chinese Dataset},
  author={Eland AI},
  year={2025},
  publisher={Hugging Face}
}