Discord-Dialogues
Discord-Dialogues is a large-scale dataset of anonymized Discord conversations for training and evaluating conversational AI models in a ChatML-friendly format.
View on Nomic Atlas
Features
- Mixed single and multi-turn exchanges
- Human-only dialogues (no bots)
- Links, embeds, and commands removed
- Filtered for ToS violations and unsafe content
- Casual Discord tone
- Two-author chains only
- Merged self-replies from the same author into a single message
- Cleaned and deduplicated for relevance
- Primarily English, with some other languages present
Use
- Fine-tuning conversational models
- Training relevance/reward models
- Dialogue generation research
Collection Policy
All data was collected adhering to Discord's Terms of Service.
Dataset Statistics (using the Hermes-3-8B tokenizer)
| Metric | Value |
|---|---|
| Samples (count) | 10,008,676 |
| Min length (tokens) | 8 |
| Max length (tokens) | 5,979 |
| Mean length (tokens) | 35.99 |
| Median length (tokens) | 30 |
| Std dev (tokens) | 21.43 |
| Total tokens | 360,279,716 |
| Total characters | 1,847,454,279 |
| Total words | 216,225,706 |
| Total Assistant Blocks | 12,775,402 |
Tokens
| Range | Count |
|---|---|
| 0–8 | 0 |
| 8–16 | 119,858 |
| 16–32 | 5,268,146 |
| 32–64 | 3,836,602 |
| 64–128 | 716,410 |
| 128–256 | 64,263 |
| 256–384 | 2,738 |
| 384–512 | 374 |
| 512–768 | 188 |
| 768–1024 | 49 |
| 1024–2048 | 37 |
| 2048–4096 | 7 |
Turns per Exchange
| Turns | Count |
|---|---|
| 2 | 7,660,871 |
| 3 | 1,532,180 |
| 4 | 496,352 |
| 5 | 173,547 |
| 6 | 75,284 |
| 7 | 33,418 |
| 8 | 16,661 |
| 9 | 8,485 |
| 10 | 4,699 |
| 11 | 2,695 |
| 12 | 1,578 |
| 13 | 979 |
| 14 | 643 |
| 15 | 403 |
| 16 | 275 |
| 17 | 178 |
| 18 | 120 |
| 19 | 87 |
| 20 | 44 |
| 21 | 46 |
| 22 | 38 |
| 23 | 28 |
| 24 | 11 |
| 25 | 10 |
| 26 | 9 |
| 27 | 9 |
| 28 | 3 |
| 29 | 8 |
| 30 | 3 |
| 31 | 1 |
| 32 | 3 |
| 33 | 3 |
| 34 | 1 |
| 37 | 2 |
| 39 | 2 |
Disclaimer
Although filtering the full data dump reduced it significantly, this dataset is still intended as a large-scale dump. For best training results, further curation to target high-signal data relevant to your goals is recommended.
License
Apache License 2.0
How to cite
@misc{discord-dialogues-2025,
title = {Discord-Dialogues},
author = {mookiezi},
year = {2025},
url={https://huggingface.co/datasets/mookiezi/Discord-Dialogues}
}