DICS Dataset Card

Dataset Details

• Dataset Type: A curated visual instruction dataset selected with DICS (Data Intrinsic Consistency Selection).
• Parent Dataset: LLaVA-1.5-665K visual instruction tuning dataset.
• Selected Scale: 166K instruction–image pairs (25% of LLaVA-1.5-665K).
• Large-Scale Corpus: DICS-6M, a six-million-sample visual instruction corpus for large-scale selection research.
• Paper: DICS: Exploring Data Intrinsic Consistency for Visual Instruction Selection (EMNLP 2026).
• Project Website: https://cqu-student.github.io/DICS/

Selection Methodology

DICS selects visual instruction data by measuring the intrinsic consistency within each image–instruction–response triplet, rather than relying solely on global diversity, heuristic filtering, or sample similarity.

It introduces two complementary sample-level signals:

• Visual Information Consistency (VIC): Measures whether visual input provides essential information for producing the response by comparing response losses with and without the image.
• Response Information Consistency (RIC): Measures whether the response is focused on and coherent with the instruction through a reverse instruction prediction task.

The final Data Intrinsic Consistency score is defined as:

$$\mathrm{D}\mathrm{I}\mathrm{C}=\min (\mathrm{V}\mathrm{I}\mathrm{C},\mathrm{R}\mathrm{I}\mathrm{C})\backslash mathrm\{DIC\}\; =\; \backslash min(\backslash mathrm\{VIC\},\; \backslash mathrm\{RIC\})$$

DICS then combines intrinsic consistency, representation diversity, and the target selection budget:

• At low selection ratios, it prioritizes high-quality samples while using farthest-point sampling to retain broad coverage.
• At high selection ratios, it preserves a high-consistency core set and removes low-quality, redundant samples.

Performance

On LLaVA-1.5-665K, training with only 25% of the original data achieves 101.40% relative performance compared with full-data fine-tuning.

Method	Data Scale	Relative Performance
Full Dataset	665K	100.00%
Random Selection	166K	96.69%
PRISM	166K	97.66%
DICS	166K	101.40%

DICS also scales effectively to the DICS-6M corpus. With InternVL3-8B, selected subsets from 15% to 85% consistently match or outperform full-data training, reaching up to 103.10% relative performance.

Intended Use

DICS is designed for:

• Efficient visual instruction tuning under constrained compute or data budgets.
• Rapid prototyping and iteration for multimodal model development.
• Research on data selection, multimodal alignment, and instruction tuning.
• Large-scale training scenarios where data quality and redundancy substantially affect training efficiency.
• Establishing compact, high-quality visual instruction subsets for fair benchmarking.

Limitations

• DICS relies on model-based scoring and representation extraction, which require non-trivial preprocessing resources.
• The method primarily targets general visual instruction tuning; highly specialized downstream tasks may benefit from task-specific selection criteria.
• Dataset availability, redistribution, and downstream use must comply with the licenses and terms of the original source datasets.

Citation

@article{hong2026dics,
  title  = {DICS: Exploring Data Intrinsic Consistency for Visual Instruction Selection},
  author = {Hong, Yuyang and Guo, Jinhui and Gu, Jiaqi and Fan, Lubin and
            Wang, Ruixiang and Ding, Kun and Wu, Yue and Xiang, Shiming and Ye, Jieping},
  year   = {2026}
}