CoolFace
Modelpublic

TorpedoSoftware/R1-Distill-Qwen-1.5B-Roblox-Luau

sourceHugging Facemitupdated 10mo agoView on Hugging Face
0likes245downloads
Model Card

R1-Distill-Qwen-1.5B-Roblox-Luau

A fine tune of deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B using boatbomber/roblox-info-dump and boatbomber/the-luau-stack for Roblox domain knowledge.

This is intended to be used for speculative decoding with boatbomber/R1-Distill-Qwen-14B-Roblox-Luau. It can be used standalone in memory constrained environments, but is not nearly as capable as the 14B model as it has so few weights that it cannot learn the same level of detail.

Recommended inference settings:

ParameterValueNotes
System PromptYou are an expert Roblox developer and Luau software engineer.Model was fine tuned with this prompt.
temperature0.5-0.7Underlying R1 Distill uses this. I've found best results with 0.55.
top_p0.95Underlying R1 Distill uses this.

Quantization done using Unsloth.

Available quants:

QuantSizeNotes
F163.56GBRetains 100% accuracy. Slow and memory hungry.
Q8_O1.89GBHigh resource use, but generally acceptable. Use when accuracy is crucial.
Q6_K1.46GBUses Q6_K for all tensors. Good for high end GPUs.
Q5KM1.29GBRecommended. Uses Q6K for half of the attention.wv and feedforward.w2 tensors, else Q5_K
Q4KM1.12GBRecommended. Uses Q6K for half of the attention.wv and feedforward.w2 tensors, else Q4_K
Q3KM0.92GBUses Q4K for the attention.wv, attention.wo, and feedforward.w2 tensors, else Q3_K. Quality is noticeably degraded.