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:

Parameter	Value	Notes
System Prompt	`You are an expert Roblox developer and Luau software engineer.`	Model was fine tuned with this prompt.
temperature	`0.5-0.7`	Underlying R1 Distill uses this. I've found best results with `0.55`.
top_p	`0.95`	Underlying R1 Distill uses this.

Quantization done using Unsloth.

Available quants:

Quant	Size	Notes
F16	3.56GB	Retains 100% accuracy. Slow and memory hungry.
Q8_O	1.89GB	High resource use, but generally acceptable. Use when accuracy is crucial.
Q6_K	1.46GB	Uses Q6_K for all tensors. Good for high end GPUs.
Q5_K_M	1.29GB	Recommended. Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q5_K
Q4_K_M	1.12GB	Recommended. Uses Q6_K for half of the attention.wv and feed_forward.w2 tensors, else Q4_K
Q3_K_M	0.92GB	Uses Q4_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else Q3_K. Quality is noticeably degraded.

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

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

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