bloomsirenix/dcxbible

Discord & Stories Dataset

A cleaned, privacy-preserving dataset combining Discord messages and creative writing for language model training.

Dataset Description

This dataset contains text from multiple sources:

• Discord message exports (sanitized for privacy)
• Creative writing and stories
• The Bible (King James Version)

Data Sources

Source	Type	Description
Discord Data Packages	Chat	Sanitized Discord messages from multiple servers
Raw Chat Exports	Chat	Additional Discord conversation data
Story Files	Creative	Original creative writing and character guides
Bible	Religious	Complete King James Version

Privacy & Sanitization

All Discord content has been aggressively sanitized to protect privacy:

Discord-Specific

• User mentions: <@123456789012345678> → @user
• Role mentions: <@&123456789012345678> → @role
• Channel mentions: <#123456789012345678> → #channel
• Custom emotes: <:name:123456789012345678> → :name:

Links & URLs (ALL REMOVED)

• HTTP/HTTPS URLs: https://... → [link]
• WWW URLs: www.example.com → [link]
• Discord invites: discord.gg/xxx → [link]
• Discord CDN: cdn.discordapp.com/... → [link]
• Media embeds: Tenor, Imgur, Giphy, YouTube, etc. → [link]
• Social media: Twitter/X, Reddit, etc. → [link]
• Email addresses: user@example.com → [email]

Personal Information

• Phone numbers: Replaced with [phone]
• Discord IDs: 18-20 digit IDs → [id]
• IP addresses: Replaced with [ip]
• Crypto addresses: ETH/BTC addresses → [crypto-address]
• API keys/tokens: Long hex strings → [hex]
• @username mentions: Generic @user

Dataset Structure

from datasets import load_from_disk

dataset = load_from_disk("hf_dataset/dataset_split")

# Access train/validation splits
train = dataset["train"]
val = dataset["validation"]

# Each example contains:
{
    "text": "The message or document content",
    "source": "Source identifier (e.g., 'discord:server:channel')",
    "type": "discord_message | story | transcript | bible_verse",
    "author": "Author name (sanitized)",
    "author_id": "Hashed user ID",
    "timestamp": "ISO timestamp (Discord only)",
    "timestamp_unix": "Unix timestamp as float",
    "message_id": "Truncated message ID",
    "channel_id": "[redacted]",
    "attachments_count": 0,
    "embeds_count": 0
}

Statistics

• Total examples: ~203K
• Train split: 182,916 (90%)
• Validation split: 20,325 (10%)
• Average text length: ~50-200 characters

By Source

Source	Examples	Type
Discord Data Packages	~178K	Chat
Raw Chat Exports	~84	Chat
Bible	~25K	Religious text
Stories/Transcripts	~103	Creative writing

Usage

Loading with Hugging Face Datasets

from datasets import load_from_disk, load_dataset

# From local directory
dataset = load_from_disk("./hf_dataset/dataset_split")

# Or from Parquet
dataset = load_dataset("parquet", data_files="hf_dataset/dataset.parquet")

For Language Model Training

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")

def tokenize_function(examples):
    return tokenizer(examples["text"], truncation=True, max_length=512)

tokenized = dataset.map(tokenize_function, batched=True)

License

• Bible text: Public Domain (King James Version)
• Discord content: Personal data, used with permission, heavily anonymized
• Stories: Original content

Citation

@dataset{discord_stories_dataset,
  title = {Discord and Stories Dataset},
  author = {Anonymous},
  year = {2024},
  url = {https://huggingface.co/datasets/USERNAME/dataset-name}
}

Limitations

• Discord content may contain informal language, slang, and typos
• Some context may be lost due to sanitization
• Not all Discord messages are high-quality prose
• Dataset is biased toward the specific servers and channels included

Contributing

To regenerate this dataset:

pip install -r requirements.txt
python convert_to_hf_dataset.py

Contact

For questions about this dataset, please open an issue on the Hugging Face Hub.