ljc-1222/LMOL-fbp5500

LMOL SCUT-FBP5500 Fold0 Checkpoint

This Hugging Face Dataset repository stores the fold0 checkpoint package and small result artifacts for the GitHub repository:

https://github.com/ljc-1222/lmol-scut-fbp5500

It is hosted as a Dataset repository to keep the GitHub repository lightweight. The package is not a redistribution of SCUT-FBP5500 images or labels.

Paper

This checkpoint is for an implementation of:

Zhenyou Liu, Xuefeng Liang, and Jian Lin, "Large Multimodal Model is a Better Comparator on Facial Beauty Prediction," ICASSP 2025. DOI: 10.1109/ICASSP49660.2025.10889943

The method uses a large multimodal model as a pairwise comparator. Given two face images, the model predicts First., Second., or Similar., and absolute SCUT-FBP5500 beauty scores are estimated from many pairwise comparisons.

Implementation Notes

This checkpoint follows the repository's default choice_token implementation. The paper writes the answer target as conditional language modeling over the natural-language answer string, but this checkpoint was trained and evaluated with a discriminative token-choice objective:

• Internal class order is First., Second., Similar..
• With the LLaVA tokenizer, each answer string is two tokens: an answer word token plus a shared period token.
• Training and pairwise evaluation use only the first answer word token (First, Second, or Similar) and do not score the period token.
• Pairwise inference is prompt-only likelihood scoring over the three answer word tokens, not free-form model.generate.
• The repository's conditional_lm branch uses the same current collator mask, so it is not a full multi-token answer-string objective unless the collator is changed.

See the GitHub repository's docs/TRAINING_ALIGNMENT.md for the detailed paper-to-code alignment notes.

Repository Layout

.
├── README.md
├── checkpoints/
│   └── fold0/
│       ├── adapter_config.json
│       ├── adapter_model.safetensors
│       ├── projector.pt
│       ├── tokenizer.json
│       ├── tokenizer_config.json
│       ├── processor_config.json
│       ├── preprocessor_config.json
│       ├── special_tokens_map.json
│       ├── chat_template.json
│       └── README.md
└── results/
    └── fold0/
        ├── metrics.json
        ├── confusion_matrix.json
        ├── progress.json
        └── README.md

Download

Run the download command from the GitHub repository root. It writes files to checkpoints/fold0/, which is the path expected by the evaluation examples.

pip install -U huggingface_hub
hf download ljc-1222/LMOL-fbp5500 \
  --repo-type dataset \
  --include "checkpoints/fold0/*" \
  --local-dir .

To download both the checkpoint and result JSON files:

hf download ljc-1222/LMOL-fbp5500 \
  --repo-type dataset \
  --include "checkpoints/fold0/*" "results/fold0/*" \
  --local-dir .

Python equivalent:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ljc-1222/LMOL-fbp5500",
    repo_type="dataset",
    allow_patterns=["checkpoints/fold0/*"],
    local_dir=".",
)

Use With The GitHub Repository

After installing dependencies, preparing SCUT-FBP5500, and building fold0 pairs, run:

python scripts/evaluate_fold.py \
  --config configs/lmol_scut_fbp5500.yaml \
  --fold-index 0 \
  --checkpoint checkpoints/fold0 \
  --batch-size 8

Full fold0 evaluation scores 374,000 target-reference pairs and can take many hours. Use --limit 256 only for a checkpoint smoke test; it does not reproduce the reported metrics.

Fold0 Result

Score metrics from results/fold0/metrics.json:

Split	PC ↑	MAE ↓	RMSE ↓
Fold 0	0.9568	0.1643	0.2130

Pairwise metrics computed from results/fold0/confusion_matrix.json and results/fold0/progress.json:

Split	Accuracy ↑	Macro Precision ↑	Macro Recall ↑	Macro F1 ↑	Weighted F1 ↑	Pairwise NLL ↓
Fold 0	0.8649	0.8145	0.8354	0.8195	0.8723	0.3384

Per-answer pairwise metrics:

Gold Label	Support	Predicted	Precision ↑	Recall ↑	F1 ↑
First.	120,836	110,502	0.9335	0.8537	0.8918
Second.	194,770	185,020	0.9557	0.9079	0.9312
Similar.	58,394	78,478	0.5541	0.7446	0.6354

Checkpoint Details

• Base model: llava-hf/llava-1.5-7b-hf
• Trainable modules: PEFT LoRA adapters on the language model and the full multimodal projector.
• Frozen modules: CLIP visual encoder.
• LoRA: r=8, lora_alpha=32, lora_dropout=0.05.
• Training objective: choice_token 3-way cross entropy over the first answer word tokens.
• Inference: prompt-only normalized likelihood over the same three answer word tokens.
• Fold: deterministic SCUT-FBP5500 fold0 with split_seed: 42.

The projector.pt file is required. Loading only the PEFT adapter is incomplete for this implementation.

Limitations And Terms

• This repository contains a checkpoint package and result metadata only. It does not contain SCUT-FBP5500 images or labels.
• Users must obtain SCUT-FBP5500 from the official source and comply with the dataset authors' terms.
• The GitHub repository currently does not specify a project license. Reuse rights should be clarified before redistribution.
• Facial beauty prediction is sensitive and subjective. Use the checkpoint only for research reproduction and with appropriate ethical review.

Citation

@inproceedings{liu2025large,
  title = {Large Multimodal Model is a Better Comparator on Facial Beauty Prediction},
  author = {Liu, Zhenyou and Liang, Xuefeng and Lin, Jian},
  booktitle = {ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages = {1--5},
  year = {2025},
  doi = {10.1109/ICASSP49660.2025.10889943}
}