Dataset Viewer

The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.

Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('json', {}), NamedSplit('validation'): ('json', {}), NamedSplit('test'): (None, {})}

Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Qwen3-VL Selector Head v1

This experiment replaces the generative JSON selector with a real five-label classifier.

Input:

<image>
Inspect this photo as a photographer.
For each dimension, decide whether there is a visually supported issue worth critiquing:
framing, camera_position, subject, focus, light_exposure.

Model:

Qwen3-VL hidden representation -> Linear(hidden_size, 5) -> masked BCEWithLogitsLoss

First run freezes the whole Qwen3-VL backbone and trains only selector_head.pt. Class imbalance is handled by capped pos_weight = negative / positive, computed only over observed labels.

Masked labels:

{
  "labels": {"framing": 1, "camera_position": 0, "subject": 1, "focus": 0, "light_exposure": 1},
  "label_mask": {"framing": 1, "camera_position": 1, "subject": 1, "focus": 1, "light_exposure": 1}
}

Native pair rows should use a full mask because the pair-derived GT checks all five dimensions.

DPC/comment rows should only supervise dimensions that are explicitly known:

{
  "labels": {"framing": 0, "camera_position": 0, "subject": 0, "focus": 1, "light_exposure": 0},
  "label_mask": {"framing": 0, "camera_position": 0, "subject": 0, "focus": 1, "light_exposure": 0}
}

This means the row contributes only focus=1 loss. Unmentioned dimensions are unknown, not negatives.

Data:

train: train_selector_native_pair_free_dedup.jsonl
source: clean native pair free labels only
DPC selector labels are allowed only with partial masks
no directed single-item expansion
no resampling

Outputs:

data_summary.json
eval_summary.json
compact_summary.json
eval_predictions.jsonl
selector_head.pt

Downloads last month: 40

Total file size:

686 MB