--- tags: - discord - chatml - conversation - dialogue - multi-turn - single-turn - fine-tuning - reward-model - llm-training - chat-dataset - open-source - anonymized-data - casual-dialogue license: apache-2.0 language: - en pretty_name: Discord-Dialogues size_categories: - 1M Discord-Dialogues

> **Discord-Dialogues** is a large-scale dataset of anonymized Discord conversations from late spring to early fall 2025 for training and evaluating realistic conversational AI models in a ChatML-friendly format. This dataset contains 7.3 million exchanges spread out over 16 million turns, with more than 139 million words. ---

discord-alpha

Nomic Atlas Map

--- ## Features - Mixed single and multi-turn exchanges - Human-only dialogues (no bots) - Filtered for ToS and harmful content - Links, embeds, and commands removed - Trading posts, code blocks, and LFG removed - 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 Use case examples: - [mookiezi/Discord-Micae-8B-Preview](https://huggingface.co/mookiezi/Discord-Micae-8B-Preview) — experimental larger model - [mookiezi/Discord-Micae-Hermes-3-3B](https://huggingface.co/mookiezi/Discord-Micae-Hermes-3-3B) — stable smaller model --- ## Filtering Pipeline This dataset was constructed with a custom multi-stage filtering toolkit: 1. **SQL filters** (`filter.sql`) Postgres regex/text filters for PII, bot/command patterns, links, embeds, and automation noise. 2. **Smart cleaner** (`smartclean.py`) Multi-stage process: normalize text, slang replacement, resample by length, and enforce structural validation. Filters out structural noise such as code blocks, trading posts, and LFG. 3. **Dedupe** (`dedupe.py`) Deduplicates conversations by hashing message chains Keeps only unique rows preferring the longest final assistant message when duplicates occur. 4. **ToS risk filter** (`tos.py`) Drops or redacts unsafe categories (sexual violence, CSA, slurs, harassment, doxxing, self-harm, extremism) and PII. Uses fuzzy/leet/diacritic-aware regex. The full filtering scripts are open source at the [dataset cleaning toolkit GitHub repository](https://github.com/mookiezi/dataset-cleaning-toolkit). --- ## Dataset Pipeline The full end-to-end pipeline is documented in the [dataset-pipeline GitHub repository](https://github.com/mookiezi/dataset-pipeline). --- ## Uncensored Edition A larger uncensored archive (10,011,633 dialogues, 215M words) is mirrored on [Internet Archive](https://archive.org/details/discord-dialogues-uncensored). --- ## Collection Policy - All data was collected adhering to Discord's [Terms of Service](https://discord.com/terms) and [Community Guidelines](https://discord.com/guidelines). --- ## Dataset Statistics (using the [NousResearch/Hermes-3-Llama-3.1-8B tokenizer](https://huggingface.co/NousResearch/Hermes-3-Llama-3.1-8B))
| Metric | Value | | ---------------------- | ------------: | | Samples (count) | 7,303,464 | | Total turns | 16,881,010 | | Total assistant turns | 9,016,287 | | Min length (tokens) | 10 | | Max length (tokens) | 2,542 | | Mean length (tokens) | 32.79 | | Median length (tokens) | 28 | | Std dev (tokens) | 16.56 | | Skew | 6.04 | | Kurtosis | 326.54 | | Total tokens | 239,458,213 | | Total characters | 1,242,238,794 | | Total words | 139,922,950 | | Avg chars per sample | 170.09 | | Avg words per sample | 19.16 | | Avg chars per word | 8.88 | | Tokens per char | 0.19 |
| Tokens | Count | | --------- | --------: | | 8–16 | 107,264 | | 16–32 | 4,278,713 | | 32–64 | 2,566,176 | | 64–128 | 334,829 | | 128–256 | 15,920 | | 256–384 | 363 | | 384–512 | 71 | | 512–768 | 78 | | 768–1024 | 30 | | 1024–2048 | 17 | | 2048–4096 | 3 |
| Turns | Count | | ----- | --------: | | 2 | 5,795,019 | | 3 | 1,038,500 | | 4 | 304,442 | | 5 | 96,758 | | 6 | 38,620 | | 7 | 15,714 | | 8 | 7,108 | | 9 | 3,391 | | 10 | 1,709 | | 11 | 909 | | 12 | 526 | | 13 | 291 | | 14 | 163 | | 15 | 113 | | 16 | 58 | | 17 | 57 | | 18 | 28 | | 19 | 20 | | 20 | 7 | | 21 | 10 | | 22 | 10 | | 23 | 2 | | 24 | 1 | | 25 | 2 | | 27 | 2 | | 29 | 1 | | 32 | 1 | | 33 | 2 |
--- ## Disclaimer Although filtering reduced the exchanges by about 75% (leaving roughly 7.5% of the full data dump), 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. In addition to the raw text, the dataset includes supporting columns such as characters, words, tokens, and turns. These provide length statistics and turn counts per exchange, which can be used for sampling, weighting, or filtering strategies. --- ## License This project is licensed under the Apache License 2.0. --- ## How to cite: ```bibtex @misc{discord-dialogues-2025, title = {Discord-Dialogues}, author = {mookiezi}, year = {2025}, url={https://huggingface.co/datasets/mookiezi/Discord-Dialogues} } ``` --- ## Related - [https://archive.org/details/discord-dialogues-uncensored](https://archive.org/details/discord-dialogues-uncensored) - [mookiezii/Discord-Hermes-3-8B](https://huggingface.co/mookiezii/Discord-Hermes-3-8B) - [mookiezi/Discord-Micae-Hermes-3-3B](https://huggingface.co/mookiezi/Discord-Micae-Hermes-3-3B) - [mookiezi/Discord-OpenMicae](https://huggingface.co/datasets/mookiezi/Discord-OpenMicae) - [NousResearch/Hermes-3-Llama-3.1-8B](https://huggingface.co/NousResearch/Hermes-3-Llama-3.1-8B) [​](https://20000.online/micae) [​](https://20000.online/openmicae) [​](https://20000.online/discord-dialogues)