Modelpublic

PortelaEmilio/calculus-ego-weights

sourceHugging Faceotherupdated 11d agoView on Hugging Face

0likes

Model Card

calculus-ego weights

Weights used by the calculus-ego pipeline and its Space. They live here, rather than in the GitHub repository, so that the Space's cold starts do not consume GitHub's LFS bandwidth.

File	What it is	License
`yolo26x-pose.pt`	Ultralytics YOLO26x-pose: person detection and 17 COCO keypoints	AGPL-3.0 (Ultralytics)
`beauty_adapter/`	LoRA (r=8) on Qwen/Qwen3.5-9B that predicts a continuous 1–10 facial attractiveness score	see below

Beauty adapter

Trained jointly on SCUT-FBP5500, CFD, MEBeauty, HotOrNot and M2B, with each dataset's scores mapped to a percentile rank on 1–10. Test-set Pearson correlation: SCUT 0.93, CFD 0.81, MEBeauty 0.81, HotOrNot 0.61, M2B 0.52.

It predicts the average rating those datasets' annotators gave, and carries their biases. The base model is Apache-2.0, but the training data were released for non-commercial research: use the adapter only under those terms.

Load it on top of the base model:

python

from peft import PeftModel
from transformers import AutoModelForImageTextToText

base = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3.5-9B", attn_implementation="eager")
model = PeftModel.from_pretrained(base, "PortelaEmilio/calculus-ego-weights", subfolder="beauty_adapter")