README.md

---
license: apache-2.0
task_categories:
- visual-question-answering
- reinforcement-learning
language:
- en
tags:
- multimodal
- regression
- deep-imbalanced-regression
- mllm
- number representation
size_categories:
- 100K<n<1M
pretty_name: MLLM Deep Imbalanced Regression Benchmarks
---

# MLLM Deep Imbalanced Regression Benchmarks

The official dataset release for **CCC-GRPO: Injecting Distributional Awareness into MLLMs via Reinforcement Learning for Deep Imbalanced Regression**, accepted at ICML 2026.

**[Yao Du](https://scholar.google.com.hk/citations?user=8krbrWsAAAAJ&hl=zh-CN), [Shanshan Song](https://scholar.google.com/citations?hl=zh-CN&user=EoNWyTcAAAAJ&view_op=list_works&sortby=pubdate&inst=1381320739207392350), [Xiaomeng Li](https://xmengli.github.io/)**

[[Paper](https://arxiv.org/abs/2605.01402)] [[Code](https://github.com/xmed-lab/CCC-GRPO)]

## Overview

This repository contains four multimodal deep imbalanced regression benchmarks used to evaluate CCC-GRPO. The training splits follow naturally long-tailed target distributions, while the test splits are balanced. No validation splits are used.

The benchmark covers 129,563 image-text regression samples across facial age estimation, movie rating prediction, and skeletal age estimation.

<p align="center">
  <img src="https://raw.githubusercontent.com/xmed-lab/CCC-GRPO/main/figures/MLLM_Numerical_3.png" width="900">
</p>

## Benchmark

| Dataset | Train | Test | Total | Target | Domain |
| --- | ---: | ---: | ---: | --- | --- |
| AgeDB-DIR | 12,208 | 2,140 | 14,348 | Age (years) | In-the-wild faces |
| IMDB-WIKI-DIR | 81,911 | 11,016 | 92,927 | Age (years) | Web-scale faces |
| IMDB-Movie-DIR | 7,049 | 1,203 | 8,252 | IMDb movie score | Movie posters |
| BoneAge-DIR | 12,528 | 1,508 | 14,036 | Bone maturity (months) | Medical imaging |
| **Total** | **113,696** | **15,867** | **129,563** | - | - |

## Repository Structure

Images and sample metadata are packaged as WebDataset tar shards:

```text
agedb/{train,test}/*.tar
imdb_wiki/{train,test}/*.tar
imdb_movie/{train,test}/*.tar
boneage/{train,test}/*.tar
```

Each WebDataset sample contains an image and a JSON record with:

| Field | Description |
| --- | --- |
| `dataset` | Benchmark name |
| `split` | `train` or `test` |
| `image` | Image filename inside the shard |
| `problem` | Regression prompt |
| `solution` | Numerical target |

The original training and test annotations are also provided:

| Dataset | Train annotation | Test annotation |
| --- | --- | --- |
| AgeDB-DIR | `agedb/agedb_train.jsonl` | `agedb/test_conversation_from_agedb.json` |
| IMDB-WIKI-DIR | `imdb_wiki/imdb_train_peak_compressed_3500_leq100.jsonl` | `imdb_wiki/test_conversation_from_imdb_leq100.json` |
| IMDB-Movie-DIR | `imdb_movie/train.jsonl` | `imdb_movie/test.json` |
| BoneAge-DIR | `boneage/boneage_train.jsonl` | `boneage/test_conversation_from_boneage.json` |

## Loading

Download the complete repository:

```bash
hf download ChanganYao/DeepImbalancedRegressionForMLLMs \
  --repo-type dataset \
  --local-dir data
```

Load one benchmark directly with `datasets`:

```python
from datasets import load_dataset

dataset = load_dataset(
    "webdataset",
    data_files={
        "train": "hf://datasets/ChanganYao/DeepImbalancedRegressionForMLLMs/agedb/train/*.tar",
        "test": "hf://datasets/ChanganYao/DeepImbalancedRegressionForMLLMs/agedb/test/*.tar",
    },
)
```

Replace `agedb` with `imdb_wiki`, `imdb_movie`, or `boneage` to load another benchmark.

## CCC-GRPO

CCC-GRPO introduces batch-level Concordance Correlation Coefficient supervision for multimodal deep imbalanced regression. Instead of assigning rewards independently to each prediction, it evaluates numerical responses in the context of the mini-batch to align the prediction and target distributions.

Training and evaluation code is available at [xmed-lab/CCC-GRPO](https://github.com/xmed-lab/CCC-GRPO).

## Citation

```bibtex
@misc{du2026injectingdistributionalawarenessmllms,
  title={Injecting Distributional Awareness into MLLMs via Reinforcement Learning for Deep Imbalanced Regression},
  author={Yao Du and Shanshan Song and Xiaomeng Li},
  year={2026},
  eprint={2605.01402},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}
```