| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - luau |
| - roblox |
| - reasoning |
| size_categories: |
| - 1K<n<10K |
| --- |
| # LuauDev-SFT-FULL 🚀 |
|
|
| THIS IS THE FULL VARIANT OF LUAUDEV. |
|
|
| This is an SFT dataset meant for training Luau(Roblox's coding language) oriented large language models. |
|
|
| These are the models which were used for data generation: |
|
|
| (no specific order) |
| - DiffusionGemma 26B A4B |
| - Deepseek v4 Flash 0731 |
| - Deepseek v4.1 flash |
| - Deepseek v4 pro |
| - Nemotron 3 Ultra 550B A55B |
| - dots3 note prev |
| - GPT OSS 120b |
| - Muse Glimmer 30B |
| - GPT OSS 20b |
| - Ling 3.0 flash |
|
|
| And more... |
|
|
|
|
|
|
| each row has exactly **5** assistant and user turns. Making it to date, the biggest Luau post training dataset ever made publicly. |
|
|
| The dataset is organized like this: |
|
|
| (columns) |
| ```conversation``` |
|
|
| Same exact format as any other swharegpt dataset, with reasoning between ```<think>, </think>``` tags. |
|
|
| Compared to luaucoder instructions v3 dataset: |
|
|
| - Higher quality data overall |
| - Better system prompt |
| - Real debugging samples⁕ |
| - Significantly improved user agent |
| |
| ⁕ The user agent did detect when the assistant generated mistakes, and did tell the assistant to fix those mistakes. (eg; bugs) |
|
|
| Compared to preview: |
| - some reasoning traces were rewritten due to them being too long and or bad to train on |
| - more rows |
| |
|
|
| Thank you for supporting and using Luaudev sft dataset! |
