Overview
Overview
TorpedoSoftware's Luau-Devstral-24B-Instruct-v0.1 is a 24 billion parameter instruction-tuned model, developed by Zack Williams (boatbomber) and funded by Torpedo Software LLC. It is a specialized fine-tune of the powerful Devstral Small 2507 model, designed to address the limited representation of Roblox game development and Luau programming in general LLM training data.
Key Capabilities
- Specialized Luau Programming: Significantly enhanced capabilities for generating and understanding Luau code, crucial for Roblox development.
- Roblox Documentation Comprehension: Improved understanding of Roblox's extensive documentation, enabling more accurate and context-aware responses.
- Efficient Local Inference: Built upon Devstral Small 2507, which is noted for strong performance in local inference scenarios.
- Custom System Prompt: Utilizes a Jinja chat template with a default system prompt to further steer the model towards Luau-specific tasks.
Training Details
The model was fine-tuned using a LoRA adapter on two custom datasets:
- The Luau Stack: 25.917 million lines (0.452 billion tokens) of real Luau code.
- Roblox Info Dump: 19.6 thousand pages (0.149 billion tokens) of multilingual Roblox documentation.
Good For
- Developers working with Roblox and Luau programming.
- Generating Luau code snippets or full scripts.
- Querying information related to Roblox development and its API.
- Applications requiring a strong understanding of the Roblox ecosystem.