Qwen2.5 Coder 1.5B Roblox (huggingface.co)

Qwen2.5 Coder 1.5B Roblox is a parameter-efficient LoRA adapter that fine-tunes Qwen2.5-Coder-1.5B-Instruct specifically for Roblox Luau development. Released with an instant browser chatbot demo and GPU-accelerated inference, it was trained on an authentic Roblox Luau corpus (90/10 train/val) filtered by code length and Luau keywords. The adapter uses LoRA (rank 8, alpha 32) applied to q_proj and v_proj, trained on TPU v5e-8 with AdamW, cosine_annealing, learning_rate 3e-5, batch_size 4 and gradient_accumulation_steps 32 for one epoch. It supports a 1,024-token context window, is Apache 2.0–licensed, and includes straightforward usage: load the Qwen base model, attach the Peft LoRA adapter, generate Luau code, and optionally merge weights for standalone deployment. For the AI/ML and game-dev communities this is a practical example of domain-specific, parameter-efficient fine-tuning: it accelerates Roblox prototyping (function generation, API usage, bug fixes, code explanation), reduces boilerplate, and standardizes team patterns while keeping model size and compute manageable. Important caveats: it’s Luau-specialized (not a general coder), may not reflect the latest Roblox API changes, and generated code should be validated in Roblox Studio. The release highlights how small adapters can deliver strong utility for niche programming ecosystems while keeping deployment lightweight.