UniFilter
Official implementation of Train a Unified Multimodal Data Quality Classifier with Synthetic Data accepted by EMNLP 2025 Findings.
Release
- [10/21/2025] 🔥 We released UniFilter model at UniFilter-Qwen3-0.6B. It is constructed on Qwen3-0.6B and SigLIP-2, which achieves better classification performance with much less model parameters.
- [10/19/2025] 🔥 We released UniFilter model at UniFilter-Qwen2.5-1.5B. Empowered by a strong 1.5B LLM backbone, the UniFilter model achieves best inference speed on quality score generation and the classification accuracy.
Introduction
UniFilter is a Unified Multimodal Data Quality Classifier for High-Quality Multimodal Data Filtering, which can generate quality scores for both image-text caption and interleaved document data. Such quality scores can be further used for high-quality data filtering to significantly strengthen the capability of pre-trained MLLMs.
This repo supports
- synthetic data generation
- UniFilter training
- quality score generation with UniFilter-Qwen2.5-1.5B.
Installation
If you just require the quality score generation, please install the customized LLaVA package only.
conda create -n unifilter python=3.10
conda activate unifilter
pip install -e LLaVA
pip install flash-attn==2.5.2 --no-build-isolation
Synthetic Data Generation for UniFilter Training
We instruct Claude-3 or Claude-3.5 to generate the desired (multimodal data example, quality score) pairs across 4 designated quality levels. The synthetic data generation scrips are:
Data Preparation for UniFilter Training
UniFilter is trained a large-scale set of (multimodal data example, quality score) pairs, which contains both caption data and interleaved document data. The synthetic multimodal example-score paired data are available at UniFilter-Post-Train-Data.
UniFilter Training
We develop the UniFilter training and scoring codebase based on LLaVA-Unified repo, which is adapted from LLaVA with the support for recent LLMs and Vision Encoders.
The architectural design of UniFilter contains three modules, the vision encoder, the visual projector, and the LLM Backbone. Different from a MLLM, the LLM Backbone does not have a language modeling head and we replace it with a score generation head. All these module parameters are specified with:
--mm_projector_type: visual projector, i.e. aapool_mlp representing average pooling vision projector with 144 tokens for one image--vision_tower: vision encoder, i.e. SigLIP-SO-400M with 384px resolution--model_name_or_path: LLM Backbone, i.e. Qwen2.5-0.5B-Instruct
Visual Projector Pre-Training (Stage 1)
Please download the 558K subset of the LLAVA-Pretrain caption dataset LLaVA-Pretrain.
Training script with DeepSpeed ZeRO-2: pretrain.sh.
UniFilter Classifier Training (Stage 2)
Training script with DeepSpeed ZeRO-3: train_classifier.sh.
Our training script will upload the metrics to wandb. The best UniFilter model is saved based on the best quality classification accuracy on the validation sets.
Quality Score Generation
Caption Data Quality Scoring
python data_scoring/data_quality_classifier_caption_scoring.py \
--model-path weizhiwang/UniFilter-Qwen2.5-1.5B \
--tar-file-path data/datacomp/medium_vanilla_filter\
--gpu-id 0 \
--batch-size 4 \
--tars-per-gpu 256 \
Interleaved Data Quality Scoring
python data_scoring/data_quality_classifier_interleaved_scoring.py \
--model-path weizhiwang/UniFilter-Qwen2.5-1.5B \
--tar-file-path data/OBELICS/obelics_webdataset\
--gpu-id 0 \
--batch-size 1 \
--tars-per-gpu 128 \
Parameters to note:
--gpu-id: for large-scale score generation using multi-machines, specify the index of machines--model-path: path to the UniFilter model checkpoint--tar-file-path: path to the webdataset image-text caption data or interleaved document data tars--tars-per-gpu: the number of webdataset tars for a single-gpu to inference on
Citation
Please cite our paper if you find this repository interesting or helpful:
@article{UniFilter,
title={Train a Unified Multimodal Data Quality Classifier with Synthetic Data},
author={Wang, Weizhi and Lin, Rongmei and Li, Shiyang and Lockard, Colin and Sarkhel, Ritesh and Lokegaonkar, Sanket and Shang, Jingbo and Yan, Xifeng and Zalmout, Nasser and Li, Xian},
journal={arXiv preprint arXiv:2510.15162},
year={2025}
}
Acknowledgement
- LLaVA: the codebase we built upon for UniFilter training.