You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

AST425 student (test25): a cheap Gemini-distilled anomaly scorer for galaxy cutouts

A 576M-parameter image scorer: Qwen3-VL-8B-Instruct's vision encoder (language model removed) with LoRA adapters merged in and a linear head on the mean image token. It was trained to reproduce Gemini's 0–50 "how scientifically interesting is this cutout?" scores (the "Argon" labels in our runs) on 40k Hubble + 40k Euclid Q1 cutouts, and scores ~1,000 images/s on one H100.

Private research model (Chris Cheng, AST425Y, University of Toronto; supervisor Jo Bovy). Not for redistribution yet.

Quick start

pip install torch transformers safetensors pillow numpy      # transformers with Qwen3-VL (tested with 5.12)
hf download TheCheng2005/ast425-student --local-dir ast425-student           # needs access to this private repo
cd ast425-student
python score_student.py examples/*.jpg                       # -> path,logit,score
from score_student import load_student, score
model = load_student(".")                      # GPU in bf16 if available, else CPU in fp32
logits, scores = score(model, ["cutout1.jpg", "cutout2.png"])

Check your setup against examples/expected_scores.csv: six Euclid Q1 test cutouts from rank 1 to the last of 89,863, with the logits from our GPU run (bf16) and from this script on a laptop CPU (fp32). Expect agreement to ~0.1 in logit.

Input and output

  • Input: any image. Each is converted to greyscale and resized to 256 × 256 (PIL bicubic), exactly as in training. The Hubble training cutouts were 150 × 150 greyscale JPEGs; Euclid's were VIS cutouts of mixed size.
  • Output: one logit per image; score = 50 · sigmoid(logit) puts it on Gemini's 0–50 scale (training minimised cross-entropy against Gemini score / 50, so the score estimates Gemini's). We use it to rank: higher = more interesting.

Files

file what
student.safetensors the full student, 1.15 GB: encoder weights (bf16, adapters merged: W + 2·B·A) + head (fp32)
config.json the encoder's architecture (Qwen3-VL-8B's vision_config), input size, base-model revision
score_student.py stand-alone loader and scorer (no peft, no other code needed)
adapter/best.pt the trained weights exactly as saved by training (PyTorch state dict: LoRA A/B + head, 8.85M numbers)
adapter/adapter.safetensors the same tensors as safetensors
examples/ six Euclid Q1 test cutouts and their expected scores

The adapter files are only needed to rebuild the student from the base model yourself: LoRA r = 16, alpha = 32 on every attn.qkv, attn.proj, linear_fc1, linear_fc2 of the vision blocks and mergers, applied to Qwen/Qwen3-VL-8B-Instruct at revision 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b; key names follow peft (visual.base_model.model.<module>.lora_A.default.weight). The three deepstack mergers' adapters are zero (the scorer never reads their output).

Training

  • Labels: Gemini 0–50 interest scores ("Argon"), one call per cutout, on 40,000 Hubble and 40,000 Euclid Q1 train cutouts.
  • Recipe: LoRA r 16 / alpha 32 / dropout 0.05 on the frozen bf16 encoder, linear head on the mean of the 64 image tokens; AdamW (no weight decay), lr 2e-4, 3% warm-up, batch 32 (16 × 2 accumulation), grad-norm clip 0.3, 2 epochs, checkpoint chosen by validation top-5% recall; one seed.

Results (held-out test splits)

test measure test25
Hubble (AnomalyMatch test split, expert-checked anomalies) share of known anomalies in the top 1 / 5 / 10% 0.63 / 0.93 / 0.95
Euclid Q1 test split (89,863) in the top 5%: lenses (of 65) / mergers (of 2,049) / problem images (of 3,458) 20 / 1,248 / 28
GZ Evo test split (never trained on) AUC for volunteers' mergers: DECaLS / DESI Legacy / CANDELS 0.90 / 0.90 / 0.87
AUC for "something odd": SDSS / UKIDSS 0.90 / 0.93

Licence and terms

  • Base model: Qwen3-VL-8B-Instruct, Apache-2.0 (© Alibaba Cloud / Qwen team). This repo contains a modified copy of its vision encoder (LoRA adapters merged in, language model removed).
  • The student was trained on Gemini outputs; check Google's terms before any public release.
  • Licence for the AST425 additions: not yet chosen (private repo).
Downloads last month
4
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for TheCheng2005/ast425-student

Finetuned
(760)
this model