asdf98/real-human-conversations

Real Human Conversations Dataset

A large-scale dataset of real human-to-human conversations collected from publicly available sources across the internet. This is not generated, synthetic, or customer-service data — it consists of actual dialogue between real people.

Dataset Overview

Property	Value
Total Conversations	2,285,855
Total Size	~1.6 GB (Parquet, zstd compressed)
Train Split	2,171,562 conversations
Validation Split	68,575 conversations
Test Split	45,718 conversations
Languages	English, Danish, Finnish, Japanese, Welsh/English, Korean
Format	Parquet

Sources

All data comes from real public human conversations on the internet:

Source	Count	Description
reddit_posts	800,000	Real Reddit discussions (title + body)
reddit_finnish	300,000	Finnish Reddit conversations
reddit_danish	299,937	Danish Reddit conversations
reddit_confessions	246,133	Real personal confessions from Reddit
opensubtitles_japanese	232,232	Real human Japanese dialogue from subtitles
opensubtitles_welsh	150,000	Welsh-English bilingual conversations
reddit_jokes	96,815	Real human humor and jokes
reddit_aita	90,000	"Am I The Asshole" real human disputes
reddit_questions	53,864	Real questions from AskReddit
korean_dialogue	16,874	Korean conversational sentences

Schema

Each row contains:

Column	Type	Description
conversation_id	string	Unique identifier
source	string	Origin dataset/source
language	string	Language code (en, fi, da, ja, cy/en, ko)
turns	int	Number of dialogue turns
text	string	Plain text with speaker labels
chatml	string	ChatML-formatted conversation
messages_json	string	JSON array of message objects
subreddit	string	Source community (if applicable)
score	int	Community score/upvotes
created_utc	string	Timestamp

Usage

from datasets import load_dataset

ds = load_dataset("asdf98/real-human-conversations")

# Access splits
train = ds["train"]
val = ds["validation"]
test = ds["test"]

# Example conversation
print(train[0]["text"])

Languages Breakdown

• English (en): ~1.29M conversations
• Danish (da): ~300K conversations
• Finnish (fi): ~300K conversations
• Japanese (ja): ~232K conversations
• Welsh/English (cy/en): ~150K conversations
• Korean (ko): ~17K conversations

License

Data is derived from publicly available sources. Individual sources retain their original licenses. Use responsibly.

Citation

@dataset{real_human_conversations_2026,
  title = {Real Human Conversations Dataset},
  author = {asdf98},
  year = {2026},
  url = {https://huggingface.co/datasets/asdf98/real-human-conversations}
}