hyeminboo/HARDBench
收藏资源简介:
--- license: cc-by-sa-4.0 configs: - config_name: default data_files: - split: train path: data/train-* - split: validation path: data/validation-* - split: test path: data/test-* - split: realx path: data/realx-* - split: vstar path: data/vstar-* dataset_info: features: - name: question dtype: string - name: answer dtype: string - name: image dtype: string - name: data_source dtype: string - name: is_unanswerable dtype: bool - name: original_question dtype: string - name: modifications dtype: string - name: pair_id dtype: float64 splits: - name: train num_bytes: 3224134 num_examples: 10000 - name: validation num_bytes: 92257 num_examples: 200 - name: test num_bytes: 238390 num_examples: 500 - name: realx num_bytes: 26869 num_examples: 82 - name: vstar num_bytes: 138203 num_examples: 382 download_size: 1456890 dataset_size: 3719853 task_categories: - visual-question-answering - image-to-text language: - en tags: - multimodal - refusal - unanswerable - visual-reasoning - tool-use - agent size_categories: - 10K<n<100K --- # HARDBench: High-resolution, systematically Altered, Refusal-Demanding Benchmark HARDBench is a benchmark and training dataset introduced in the paper **"Rotus: Calibrated Refusal Optimization for Multimodal Tool-Use Agents"**. It is designed to train and evaluate **refusal calibration** in multimodal tool-use agents — specifically, whether agents can appropriately abstain from answering when visual evidence is insufficient, rather than generating overconfident or hallucinated responses. 📄 Paper (coming soon) | 💻 [GitHub](https://github.com/ewha-mmai/rotus) --- ## Overview State-of-the-art multimodal agents finetuned with reinforcement learning often lose the ability to refuse unanswerable questions — even when their own tool outputs fail to provide sufficient evidence. HARDBench addresses this gap by providing systematically constructed unanswerable questions paired with real, high-complexity visual scenarios. Unlike prior benchmarks that rely on artificially manipulated images or narrow visual domains, HARDBench generates unanswerability by modifying **semantic components** of real questions — specifically **Objects, Relations, and Attributes (ORA)** — through a structured three-step pipeline. --- ## Dataset Splits and Composition | Split | Source | # Samples | |-------|--------|-----------| | Train | PixMo Counting, TallyQA, ArxivQA, MM-adaptive-CoF RL | 10,000 | | Test (In-domain) | PixMo Counting, TallyQA, ArxivQA, MM-adaptive-CoF RL | 500 | | Test (HARDBench-RealX) | RealX-Bench | 82 | | Test (HARDBench-V*) | V*Bench | 382 | - Training set maintains a **3:1 ratio** of answerable to unanswerable questions. - Test sets contain equal distributions of answerable and unanswerable questions. --- ## Data Sources & Licenses HARDBench questions reference images from the following source datasets. Images must be downloaded separately: | Dataset | Link | License | |---------|------|---------| | ArxivQA | [MMInstruction/ArxivQA](https://huggingface.co/datasets/MMInstruction/ArxivQA) | CC-BY-SA-4.0 | | MM-adaptive-CoF RL | [xintongzhang/CoF-RL-Data](https://huggingface.co/datasets/xintongzhang/CoF-RL-Data) | SA-1B Dataset Research License | | PixMo-Count | [allenai/pixmo-count](https://huggingface.co/datasets/allenai/pixmo-count) | ODC-BY-1.0 | | TallyQA | [tallyqa.zip](https://github.com/manoja328/tallyqa) | Apache License 2.0 | | RealXBench | [glowol/RealXBench](https://huggingface.co/datasets/glowol/RealXBench) | Apache License 2.0 | | V*Bench | [craigwu/vstar_bench](https://huggingface.co/datasets/craigwu/vstar_bench) | SA-1B Dataset Research License | --- ## Citation ```bibtex coming soon ```





