tellang/yeji-fortune-telling-ko-v9

Yeji Fortune-Telling Korean SFT Dataset v9

Dataset Summary

Yeji-fortune-telling-ko-v9 is a comprehensive supervised fine-tuning (SFT) dataset for training Korean language models on fortune-telling and divination tasks. It contains 31,625 high-quality instruction-output pairs across four divination domains: Astrology, Bazi, Tarot, and Hwatu.

Dataset Statistics

Total Records: 31,625

Domain	Records	Percentage
Astrology (점성술)	10,000	31.6%
Bazi (사주)	8,423	26.6%
Tarot (타로)	6,621	20.9%
Hwatu (화투)	6,581	20.8%

Data Distribution by Source

Source	Records	Percentage
horoscope_saved	9,844	31.1%
gpt5mini_synthesized	8,423	26.6%
gemini-3-flash	6,621	20.9%
hwatu_synthesized	6,581	20.8%
horoscope_saved_gpt5mini_fixed	156	0.5%

File Descriptions

1. train_alpaca.jsonl

Format: Alpaca instruction-following format

Each line contains:

• instruction: User's fortune-telling question or request
• input: Additional context (mostly empty)
• output: Expert-level divination interpretation and advice
• domain: Divination type (astrology, bazi, tarot, hwatu)
• source: Data origin

Best For: Traditional instruction-tuning, models expecting instruction/input/output structure

2. train_chatml.jsonl

Format: ChatML conversation format (Qwen, Llama Chat compatible)

Each line contains:

• messages: List of user and assistant turn messages
• domain: Divination type

Best For: Multi-turn conversation fine-tuning, chat-based applications

3. yeji_v9_train.jsonl

Legacy format (Alpaca-based). Use train_alpaca.jsonl for new projects.

Domains

Astrology (점성술) - 10,000 records

• Western zodiac horoscope readings
• Personality, relationships, career, health predictions
• Astrological aspects and planetary influences

Bazi (사주) - 8,423 records

• Chinese Four Pillars of Destiny analysis
• Fortune based on birth time and date
• Career, relationships, personality insights

Tarot (타로) - 6,621 records

• Celtic Cross and other spread interpretations
• Card meanings and combined symbolism
• Life guidance and decision-making support

Hwatu (화투) - 6,581 records (NEW in v9)

• Traditional Korean card divination
• Card-based decision-making guidance
• Yes/No and choice-based fortunes

Usage Examples

Loading with HuggingFace Datasets

from datasets import load_dataset

dataset = load_dataset("json", data_files="train_alpaca.jsonl", split="train")
print(f"Total records: {len(dataset)}")
print(dataset[0])

Filtering by Domain

import json

tarot_records = [
    json.loads(line) for line in open("train_alpaca.jsonl")
    if json.loads(line)["domain"] == "tarot"
]
print(f"Tarot records: {len(tarot_records)}")

Fine-tuning with LoRA

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer
from peft import LoraConfig, get_peft_model

dataset = load_dataset("json", data_files="train_alpaca.jsonl", split="train")

def format_alpaca(example):
    return {
        "text": f"[INST] {example['instruction']} [/INST] {example['output']}"
    }

dataset = dataset.map(format_alpaca)

model_name = "Qwen/Qwen2-7B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

lora_config = LoraConfig(
    r=8,
    lora_alpha=32,
    lora_dropout=0.05,
    task_type="CAUSAL_LM"
)
model = get_peft_model(model, lora_config)

args = TrainingArguments(
    output_dir="./results",
    num_train_epochs=3,
    per_device_train_batch_size=4,
)

trainer = Trainer(model=model, args=args, train_dataset=dataset)
trainer.train()

Dataset Characteristics

Quality Assurance

All records validated for:

• JSON integrity and valid parsing
• Required fields present (instruction, output, domain, source)
• Proper domain categorization
• Native Korean language quality
• Semantic coherence of instruction-output pairs
• Duplication removal

Quality Features

• Professional expertise in divination domains
• Cultural authenticity and sensitivity
• Balanced compassionate and practical tone
• Detailed, long-form interpretations
• Multi-perspective divination approaches

Version History

v9 (2026-02-05) - Current Release

New:

• Added Hwatu (화투) domain: 6,581 records
• Total expansion: 25,044 → 31,625 records
• Both Alpaca and ChatML formats
• Enhanced validation pipeline

Composition:

• Astrology: 10,000 records
• Bazi: 8,423 records
• Tarot: 6,621 records
• Hwatu: 6,581 records (NEW)

v8 (Previous)

• 25,044 records
• Three domains: Astrology, Bazi, Tarot
• Alpaca format only

License & Citation

License: CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0)

Non-commercial use only. Attribution required.

Citation:

@dataset{yeji_fortune_telling_v9_2026,
  title={Yeji Fortune-Telling Korean SFT Dataset v9},
  author={SSAFY Fortune-Telling Project},
  year={2026},
  month={February},
  publisher={HuggingFace Datasets},
  url={https://huggingface.co/datasets/tellang/yeji-fortune-telling-ko-v9},
  note={31,625 Korean fortune-telling instruction-output pairs}
}

Tags

• Language: Korean (한국어)
• Tasks: text-generation, instruction-following, conversation, fine-tuning
• Domains: divination, fortune-telling, astrology, bazi, tarot, hwatu, culture
• Models: qwen, llama, causal-lm, sft

Important Considerations

Content Notes

• Traditional divination content (not scientific)
• For cultural and entertainment purposes
• Responses are interpretive, not factual predictions
• Not for medical, financial, or critical decision-making

Recommendations

1. Cross-validate outputs with domain experts
2. Consider domain-specific LoRA adapters
3. Implement safety filters for factual claims
4. Be aware of cultural assumptions

Metadata

Property	Value
Dataset Size	115 MB (combined JSONLs)
Total Examples	31,625
Language	Korean (ko)
License	CC-BY-NC-4.0
Last Updated	2026-02-05
Status	Active

---

Identifier: tellang/yeji-fortune-telling-ko-v9
Release Date: February 5, 2026
Card Version: 1.0