Configuration Parsing Warning:In adapter_config.json: "peft.task_type" must be a string

Glimmer for Rangrez: first-dip colourist (LoRA on Muse Glimmer 30B)

Given a target colour in CIELAB, proposes the first dye recipe (turquoise, fuchsia and lemon yellow reactive dyes, g/L) for the Rangrez self-driving dye lab. Fine-tuned with LoRA (rank 8, 2 epochs; trained on Modal, 1x H100, PEFT) on 2,900 of the lab's recorded dips: each the recipe dipped and the colour SAM 3.1 measured on the dried cloth in a real-looking phone photo. Dataset: TechTush/rangrez-lab-dips.

The dips come from Rangrez's simulator (a hidden dye lot plus realistic photos), not from a physical dye house, and the adapter knows that lot.

First-dip error on 100 held-out targets (CIEDE2000 of the colour the recipe really dyes)

Method Median p90 Under 3
supplier model (today's first dip) 4.1 8.52 38%
model refitted to 400 of the lab's dips 2.05 3.56 78%
Muse Glimmer 30B, base 11.38 19.4 8%
Muse Glimmer 30B, tuned 0.92 1.76 96%

Code: https://github.com/TusharTechs/rangrez

Downloads last month
1
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for TechTush/rangrez-glimmer-lora

Adapter
(47)
this model

Space using TechTush/rangrez-glimmer-lora 1