Implementation:Hiyouga LLaMA Factory Dataset Info Registry

Knowledge Sources	Hiyouga_LLaMA_Factory
Domains	NLP, Training_Data
Last Updated	2026-02-06 19:00 GMT

Overview

dataset_info.json is the central dataset registry that maps dataset names to their file locations, data formats, column configurations, and hub URLs, serving as the configuration backbone for all data loading in LLaMA Factory.

Description

The file is a JSON dictionary where each key represents a dataset name (e.g., "alpaca_en_demo", "dpo_en_demo") and the value is an object describing how to locate and parse that dataset. Properties include file_name for local files, hf_hub_url/ms_hub_url/om_hub_url for remote hub datasets, formatting to specify the data format ("alpaca" by default or "sharegpt"), columns for field name mappings, tags for role/content tag overrides, and ranking to flag preference/DPO datasets.

The data parser module (src/llamafactory/data/parser.py) reads this file at startup to discover all available datasets and determine how to load, format, and process each one during training.

Usage

Users extending LLaMA Factory with custom datasets add entries to this file, specifying at minimum a file_name or hub URL. The training pipeline references datasets by their registry key name. This file supports both local demo files (shipped with the repository) and remote datasets hosted on Hugging Face Hub, ModelScope, or OpenMind Hub.

Code Reference

Source Location

• Repository: Hiyouga_LLaMA_Factory
• File: data/dataset_info.json

Data Format

{
  "alpaca_en_demo": {
    "file_name": "alpaca_en_demo.json"
  },
  "glaive_toolcall_en_demo": {
    "file_name": "glaive_toolcall_en_demo.json",
    "formatting": "sharegpt",
    "columns": {
      "messages": "conversations",
      "tools": "tools"
    }
  },
  "dpo_en_demo": {
    "file_name": "dpo_en_demo.json",
    "ranking": true,
    "formatting": "sharegpt",
    "columns": {
      "messages": "conversations",
      "chosen": "chosen",
      "rejected": "rejected"
    }
  },
  "kto_en_demo": {
    "file_name": "kto_en_demo.json",
    "formatting": "sharegpt",
    "columns": {
      "messages": "messages",
      "kto_tag": "label"
    },
    "tags": {
      "role_tag": "role",
      "content_tag": "content",
      "user_tag": "user",
      "assistant_tag": "assistant"
    }
  },
  "alpaca_en": {
    "hf_hub_url": "llamafactory/alpaca_en",
    "ms_hub_url": "llamafactory/alpaca_en",
    "om_hub_url": "HaM/alpaca_en"
  }
}

I/O Contract

Schema

Field	Type	Required	Description
file_name	string	No	Local file path relative to the data/ directory
hf_hub_url	string	No	Hugging Face Hub dataset identifier
ms_hub_url	string	No	ModelScope Hub dataset identifier
om_hub_url	string	No	OpenMind Hub dataset identifier
formatting	string	No	Data format: "alpaca" (default) or "sharegpt"
ranking	boolean	No	Whether this is a preference/ranking dataset (for DPO/RM training)
subset	string	No	Dataset subset name for hub datasets
split	string	No	Dataset split to use (e.g., "train", "validation")
columns	object	No	Mapping of internal column names to actual column names in the data
tags	object	No	Mapping of role/content tag names for ShareGPT-format datasets

Column Mapping Options

Internal Name	Description	Used In
prompt	Instruction/question field	Alpaca format
response	Output/answer field	Alpaca format
system	System prompt field	Alpaca format
messages	Conversation messages list	ShareGPT format
chosen	Preferred response	DPO/ranking datasets
rejected	Dispreferred response	DPO/ranking datasets
kto_tag	Binary quality label	KTO datasets
tools	Tool/function definitions	Tool-calling datasets
images	Image file paths	Multimodal datasets
videos	Video file paths	Multimodal datasets
audios	Audio file paths	Multimodal datasets

Usage Examples

# Adding a custom dataset to dataset_info.json
# Append to the JSON object in data/dataset_info.json:
#
# "my_custom_dataset": {
#     "file_name": "my_custom_dataset.json",
#     "formatting": "alpaca",
#     "columns": {
#         "prompt": "question",
#         "response": "answer"
#     }
# }

# Then reference it in training:
# llamafactory-cli train \
#     --dataset my_custom_dataset \
#     --stage sft \
#     --model_name_or_path meta-llama/Llama-2-7b-hf \
#     --output_dir output/custom_sft

# Programmatic inspection of the registry
import json

with open("data/dataset_info.json", "r", encoding="utf-8") as f:
    registry = json.load(f)

print(f"Total registered datasets: {len(registry)}")
local_datasets = {k: v for k, v in registry.items() if "file_name" in v}
hub_datasets = {k: v for k, v in registry.items() if "hf_hub_url" in v}
ranking_datasets = {k: v for k, v in registry.items() if v.get("ranking")}
print(f"Local datasets: {len(local_datasets)}")
print(f"Hub datasets: {len(hub_datasets)}")
print(f"Ranking/DPO datasets: {len(ranking_datasets)}")

Related Pages

• Hiyouga_LLaMA_Factory_Alpaca_En_Demo_Data - English Alpaca demo dataset registered here
• Hiyouga_LLaMA_Factory_Alpaca_Zh_Demo_Data - Chinese Alpaca demo dataset registered here
• Hiyouga_LLaMA_Factory_DPO_En_Demo_Data - English DPO demo dataset registered here
• Hiyouga_LLaMA_Factory_DPO_Zh_Demo_Data - Chinese DPO demo dataset registered here
• Hiyouga_LLaMA_Factory_Glaive_Toolcall_En_Demo_Data - English tool-calling demo dataset registered here
• Hiyouga_LLaMA_Factory_Glaive_Toolcall_Zh_Demo_Data - Chinese tool-calling demo dataset registered here
• Hiyouga_LLaMA_Factory_Identity_Data - Identity training data registered here
• Hiyouga_LLaMA_Factory_KTO_En_Demo_Data - English KTO demo dataset registered here