| --- |
| license: apache-2.0 |
| language: |
| - en |
| tags: |
| - discord |
| - chatml |
| - conversation |
| - dialogue |
| - multi-turn |
| - single-turn |
| - luna-protocol |
| size_categories: |
| - 100M<n<1B |
| pretty_name: Luna Protocol Dataset |
| --- |
| |
| > **Discord-Dialogues-Preprocessed-Luna-Protocol** is a preprocessed fork of [mookiezi/Discord-Dialogues](https://huggingface.co/datasets/mookiezi/Discord-Dialogues), adapted for fine-tuning Qwen2.5-family models as part of the **Luna Protocol** project. |
|
|
| This dataset contains anonymized Discord conversations for training and evaluating realistic conversational AI models in a ChatML-friendly format. It is derived directly from `mookiezi/Discord-Dialogues` with two targeted preprocessing steps applied (see below) — the underlying conversations, filtering pipeline (PII removal, ToS-risk filtering, dedup), and licensing are unchanged from the original. |
|
|
| ## What was changed from the original dataset |
|
|
| The original `mookiezi/Discord-Dialogues` ships each example as a ChatML string terminated with the literal token `<|end_of_text|>`. That token is **not** a recognized special token in the Qwen2.5 tokenizer (it originates from Llama 3 / GPT-oss style tokenizers), which meant it would get split into several ordinary sub-tokens during fine-tuning instead of being treated as a single end-of-sequence marker — actively harmful for teaching a model where to stop generating. |
|
|
| Two preprocessing steps were applied on top of the original `text` column: |
|
|
| 1. **EOS token fix** — every occurrence of `<|end_of_text|>` was replaced with the actual EOS token of the target tokenizer (`<|im_end|>` for Qwen2.5). Any example that didn't already end with an EOS marker had one appended. |
| 2. **Length filtering** — examples were filtered using the dataset's own `tokens` column, keeping only exchanges with **8 to 512 tokens**. This removes degenerate very-short exchanges and caps sequence length for training efficiency. Only 127 out of 7,300,966 examples fell outside this range and were dropped. |
|
|
| No other changes were made: no re-cleaning, no re-deduplication, no additional filtering of content. The `tokens`, `turns`, `characters`, and `words` metadata columns from the original dataset are preserved as-is (note: `tokens` reflects the original tokenizer's count, not a recount after the EOS substitution). |
|
|
| ## Columns |
|
|
| | Column | Type | Description | |
| | ------------ | ------ | ----------------------------------------------------------------------------- | |
| | `text` | string | ChatML-formatted conversation, EOS-corrected for Qwen2.5 (`<|im_start|>`/`<|im_end|>`) | |
| | `tokens` | int64 | Token count from the original dataset (used for the 8–512 filter) | |
| | `turns` | int64 | Number of conversational turns | |
| | `characters` | int64 | Character count | |
| | `words` | int64 | Word count | |
|
|
| ## Intended use |
|
|
| Fine-tuning small-to-mid-size causal language models (tested with Qwen2.5-1.5B-Instruct via QLoRA/Unsloth) to emulate informal, human Discord conversation style, as part of the **Luna Protocol** model family (e.g. `Luna-Protocol-1.5B-Discord-Dialogues`). |
|
|
| If you're targeting a different base model whose EOS token differs from Qwen2.5's `<|im_end|>`, you'll want to re-run the EOS substitution step against the original dataset rather than reusing this fork as-is. |
|
|
| ## Source dataset |
|
|
| - **Original dataset**: [mookiezi/Discord-Dialogues](https://huggingface.co/datasets/mookiezi/Discord-Dialogues) (7.3M exchanges, 16M turns, 139M+ words, anonymized, English, late spring–early fall 2025) |
| - **License**: Apache License 2.0 (inherited from the original dataset) |
|
|
| ## Citation |
|
|
| If you use this dataset, please cite the original: |
|
|
| ``` |
| @misc{discord-dialogues-2025, |
| title = {Discord-Dialogues}, |
| author = {mookiezi}, |
| year = {2025}, |
| url = {https://huggingface.co/datasets/mookiezi/Discord-Dialogues} |
| } |
| ``` |
