README.md

---
pretty_name: "FIKA-Bench"
license: other
language:
  - zh
  - en
task_categories:
  - visual-question-answering
tags:
  - image
  - fine-grained-recognition
  - fine-grained-knowledge
  - agent-evaluation
size_categories:
  - n<1K
gated: true
extra_gated_heading: "Request access to FIKA-Bench"
extra_gated_description: "Access is manually reviewed. Please provide your real name, affiliation, and a verifiable institutional email address."
extra_gated_button_content: "Submit access request"
extra_gated_prompt: >
  FIKA-Bench is released for non-commercial research and benchmark evaluation.
  By requesting access, you agree not to redistribute the dataset, annotations,
  answers, evidence links, or substantially equivalent derived files on the
  Internet or other public platforms. This restriction is important for
  preserving fair agent and model evaluation, because public copies of the
  ground-truth answers or evidence could allow future systems to retrieve the
  benchmark directly and produce unfair results.
extra_gated_fields:
  "Full name": text
  "Affiliation / institution": text
  "Verifiable institutional email": text
  "Intended use":
    type: select
    options:
      - "Academic research"
      - "Benchmark reproduction"
      - "Non-commercial model evaluation"
      - label: "Other non-commercial research"
        value: "other_noncommercial_research"
  "I agree not to redistribute FIKA-Bench files, annotations, answers, evidence links, or substantially equivalent derived files on the Internet or other public platforms": checkbox
  "I agree not to use the ground-truth answers, fine-grained labels, evidence links, or annotations as training data, retrieval data, prompt context, memory, or tool-accessible resources for models evaluated on FIKA-Bench": checkbox
  "I agree to follow the FIKA-Bench Research and Evaluation License": checkbox
---

# FIKA-Bench: From Fine-grained Recognition to Fine-Grained Knowledge Acquisition

FIKA-Bench is a fine-grained visual knowledge acquisition benchmark for
evaluating whether multimodal models and agentic systems can go beyond
recognizing a fine-grained visual entity and answer knowledge-intensive
questions about it. The benchmark contains image-question-answer samples across
real-life and public-image sources, with bilingual category taxonomy and
evidence supporting each gold answer.


## Files

This Hugging Face repository distributes the benchmark as a gated archive:

```text
README.md
LICENSE
fika-bench-testset.zip
fika-bench-testset.sha256
```

After extracting `fika-bench-testset.zip`, the benchmark files have the
following structure:

```text
data/
├── README.md
├── manifest.jsonl
├── agent_manifest.jsonl
├── taxonomy.json
├── taxonomy.tsv
├── checksums.sha256
└── images/
```

- `manifest.jsonl`: canonical test manifest for standard local/API evaluation.
- `agent_manifest.jsonl`: derived compatibility manifest for OpenClaw/OpenCode-style evaluation.
- `images/`: redacted test images with EXIF metadata removed.
- `taxonomy.json` / `taxonomy.tsv`: bilingual category taxonomy.
- `checksums.sha256`: checksums for files in this package.

In `manifest.jsonl`, `answer` is the canonical gold answer used for evaluation.
`fine_label` stores the fine-grained annotation label associated with the
sample; in the current version, it matches `answer` after normal whitespace
trimming. `evidence` records a supporting source for the gold answer.

## Download and Use

After your access request is approved, download the archive with the Hugging
Face CLI:

```bash
hf auth login
hf download oking0197/FIKA-Bench --repo-type dataset --local-dir ../release
```

For reproduction with the official code repository, place or download
`fika-bench-testset.zip` under `../release/` relative to the cloned repository:

```text
/path/to/workdir/
├── release/
│   └── fika-bench-testset.zip
└── fika-bench/
```

The evaluation scripts automatically extract the archive when `data/manifest.jsonl`
is missing. You can also extract it manually from the repository root:

```bash
unzip ../release/fika-bench-testset.zip
```

## Dataset Statistics

- Total samples: 311.
- Real-life samples: 116.
- Public-source samples: 195.
- Top-level categories: Product, Nature, Transport, and Culture.

## License

The benchmark package is released under the FIKA-Bench Research and Evaluation
License included with this dataset. The license applies to the FIKA-Bench
benchmark package and annotations, including questions, answers, fine-grained
labels, evidence links, taxonomy metadata, manifests, and benchmark
organization.

Images and other third-party materials remain subject to their original
licenses, terms, and rights. This dataset license does not supersede any
third-party license that applies to an individual image or source item.

## Citation

If you use FIKA-Bench, please cite:

```bibtex
@misc{li2026fikabenchfinegrainedrecognitionfinegrained,
      title={FIKA-Bench: From Fine-grained Recognition to Fine-Grained Knowledge Acquisition},
      author={Geng Li and Yuxin Peng},
      year={2026},
      eprint={2605.13193},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.13193},
}
```