A Roblox Luau focused finetune of
Devstral Small 2507.
Devstral Small 2507 is a powerful choice for local inference, achieving SOTA open source results at just 24B parameters. However, Roblox gamedev and Luau programming are generally not well represented in LLM training data. This model fine tunes Devstral on a corpus of permissively licensed Luau code and Roblox documentation, improving the model's Luau programming capabilities. Additionally, the jinja chat template contains a default system prompt that steers the model's Luau capabilities even further.
-
25.917M lines of real Luau code, 0.452B tokens. Format:
1Repository: {repo_name}
2Repository Description: {repo_description}
3
4File Path: `{file_path}`
5File Content:
6```Lua
7{file_content}
8```\
-
19.6K pages of multilingual Roblox documentation, 0.149B tokens. Format:
1Roblox Creator Docs: {url}
2```md
3{content}
4```\
Trained a LoRA adapter (r=64) at full precision on two epochs of the dataset for a total of 54,630 steps and 43.40 E FLOPs. Then merged the final adapter checkpoint into a BF16 model.
The imatrix for the GGUF quantizations was computed using 5.73MB of text containing a combination of
technical.txt,
groups_merged.txt, and content from
the-luau-stack &
roblox-info-dump. This created an imatrix that is well suited to the specialized tasks this model is designed for while still maintaining broader intelligence as well. While we do provide several quantizations already, the
imatrix.gguf is included in this repository should you want to create other quants yourself.
