docs: new dataset card for r3b_v1 layout

README.md

@@ -1,86 +1,116 @@
 ---
-license: cc-by-4.0
-pretty_name: R3-Bench Research Data
+license: apache-2.0
+task_categories:
+  - image-to-text
+  - visual-question-answering
+language:
+  - en
 size_categories:
-- n>1T
+  - n<1K
 tags:
-- benchmark
-- vision-language
-- video-understanding
-- evaluation
+  - benchmark
+  - reflection
+  - rectification
+  - text-to-image
+  - visual-reasoning
+  - image-editing
+pretty_name: R3-Bench
 ---
 
-# R3-Bench Research Data
+# R3-Bench: Reason-Reflect-Rectify Benchmark
 
-> **⚠️ This is not a HuggingFace `datasets` library dataset.**
-> This repo is a raw collection of project directories from R3-Bench research,
-> packaged as gzip'd tar files. Use the download + extraction instructions
-> below — do **not** use `datasets.load_dataset()` on this repo.
+**Code**: [github.com/xiaomoguhz/R3-Bench](https://github.com/xiaomoguhz/R3-Bench)
+**Paper**: *Benchmarking and Evolving Reason-Reflect-Rectify for Reflective Visual Generation* (ICML 2026, under review).
 
-## What's inside
+R³-Bench evaluates **reflective visual generation**: given a generated image and the original prompt, a model must (i) reason about whether the image matches the prompt, (ii) explain any discrepancy, and (iii) propose a rectification action. The benchmark measures two complementary scores:
 
-Large binary outputs from the R3-Bench evaluation framework research — roughly **323 GB** across multiple tarballs. Each file corresponds one-to-one to a directory under the source working tree:
+- **S_ref** — Reflective Verdict Score (verdict + explanation correctness, judged by an LLM)
+- **S_rect** — Rectification Score (normalised VQA-alignment gain after applying the proposed edit)
 
-| File | Source directory | Size | Shards |
-|---|---|---|---|
-| `alive.tar.gz` | `alive/` | 10.7 kB | 1 |
-| `backup_bin.tar.gz.part_{aa..ae}` | `backup_bin/` | ~193 GB | 5 |
-| `compare_methods_OmniVerifier_data.tar.gz` | `compare_methods/OmniVerifier/` | 2.09 GB | 1 |
-| `compare_methods_Step1X-Edit.tar.gz` | `compare_methods/Step1X-Edit/` | 6.23 GB | 1 |
-| `compare_methods_Step1X-Editv2.tar.gz` | `compare_methods/Step1X-Editv2/` | 129 MB | 1 |
-| `elo_human_eval.tar.gz` | `elo_human_eval/` | 146 MB | 1 |
-| `eval.tar.gz` | `eval/` | 3.02 GB | 1 |
-| `exps.tar.gz.part_{aa..ac}` | `exps/` | ~103 GB | 3 |
-| `images.tar.gz` | `images/` | 725 MB | 1 |
-| `iterative_ablation_output.tar.gz` | `iterative_ablation_output/` | 1.94 GB | 1 |
-| `logs.tar.gz` | `logs/` | 121 kB | 1 |
-| `output.tar.gz` | `output/` | 6.58 GB | 1 |
-| `paper_case.tar.gz` | `paper_case/` | 5.62 GB | 1 |
+## Contents
 
-Files larger than the HuggingFace single-file LFS limit were split into ~50 GB shards with `split -b 45G`.
+| File | Description |
+|---|---|
+| `images.tar.gz` | 670 source images (PNG, ~717 MB), organised as `images/{category}/{verdict}/r3b_{idx:06d}.png` |
 
-## Download and extract
+The 670-sample **prompt + ground-truth JSONL** ships with the code repository at [`r3bench/data/r3bench.jsonl`](https://github.com/xiaomoguhz/R3-Bench/blob/main/r3bench/data/r3bench.jsonl) — labels travel with the eval code so version drift is captured by git.
 
-```bash
-# 1. Install the HF client
-pip install -U huggingface_hub
-
-# 2. Download everything to a local directory
-huggingface-cli download xiaomoguhzz/R3-Bench-data \
-    --repo-type dataset \
-    --local-dir R3-Bench-data/
-cd R3-Bench-data/
-
-# 3. Merge split shards back into monolithic tarballs
-cat backup_bin.tar.gz.part_* > backup_bin.tar.gz
-cat exps.tar.gz.part_*        > exps.tar.gz
-rm backup_bin.tar.gz.part_* exps.tar.gz.part_*
-
-# 4. Extract each tarball
-for f in *.tar.gz; do
-    echo "extracting $f ..."
-    tar xzf "$f"
-done
-```
+## Schema
 
-After extraction the source working tree layout is reproduced:
+Each JSONL record:
 
-```
-alive/   backup_bin/   compare_methods/   elo_human_eval/   eval/
-exps/    images/       iterative_ablation_output/   logs/
-output/  paper_case/
+```json
+{
+  "idx": 0,
+  "original_prompt": "a black candle and a white holder",
+  "bad_image": "images/color/false/r3b_000000.png",
+  "answer": false,
+  "explanation": "The white object is a candle, not a holder as specified in the prompt.",
+  "category": "color",
+  "generated_qa": {
+    "yn_question_list": ["Is there a candle in the image?", "..."]
+  }
+}
 ```
 
-## Why the Dataset Viewer shows an error
+| Field | Type | Description |
+|---|---|---|
+| `idx` | int | Unique sample id (0–669) |
+| `original_prompt` | str | The text-to-image prompt |
+| `bad_image` | str | Image path relative to the data root |
+| `answer` | bool | Ground-truth verdict: does the image match the prompt? (`false` = mismatch) |
+| `explanation` | str | Ground-truth discrepancy description (used as S_ref reference) |
+| `category` | str | Error dimension: color · object · numeracy · spatial · shape · texture · complex · non |
+| `generated_qa.yn_question_list` | list[str] | Yes/no VQA probes used by the S_rect rectification scorer |
+
+## Category & verdict distribution (670 total)
+
+| Category | false (mismatch) | true (match) | Total |
+|---|---:|---:|---:|
+| color    | 71 | 26 | 97 |
+| complex  | 46 | 48 | 94 |
+| non      | 7  | 42 | 49 |
+| numeracy | 73 | 23 | 96 |
+| object   | 54 | 17 | 71 |
+| shape    | 72 | 20 | 92 |
+| spatial  | 76 | 25 | 101 |
+| texture  | 49 | 21 | 70 |
+| **Sum**  | **448** | **222** | **670** |
+
+`non` = "no error" probe samples (image matches prompt) used to calibrate false-positive reflection.
+
+## Usage
 
-The HuggingFace Dataset Viewer auto-detects tar files and tries to parse them as [WebDataset format](https://huggingface.co/docs/hub/datasets-webdataset). This repo is a raw tarball collection (not WebDataset), so the viewer reports a `SplitsNotFoundError`.
+```bash
+# 1. Download images
+hf download xiaomoguhzz/R3-Bench-data images.tar.gz \
+    --repo-type dataset --local-dir /path/to/r3bench-data
+cd /path/to/r3bench-data && tar -xzf images.tar.gz
+# → /path/to/r3bench-data/images/{category}/{true,false}/r3b_{idx:06d}.png
+
+# 2. Run the 4-step pipeline (clone the code repo first)
+git clone https://github.com/xiaomoguhz/R3-Bench && cd R3-Bench
+export R3BENCH_DATA_DIR=/path/to/r3bench-data
+
+bash scripts/run_reflection.sh 8 qwen2.5vl    # Step 1 — reflection
+bash scripts/run_edit.sh       8 qwen2.5vl qwen_image_2511  # Step 2 — editing
+bash scripts/eval_reflection.sh qwen2.5vl       # Step 3 — S_ref
+bash scripts/eval_edit.sh      qwen2.5vl qwen_image_2511    # Step 4 — S_rect
+```
 
-**This is expected** and does not affect downloads — the error only concerns the preview panel, not the files themselves.
+See the [code repo README](https://github.com/xiaomoguhz/R3-Bench) for full pipeline details, supported backends, and how to plug in a custom reflection / editor model.
 
-## Related code
+## Citation
 
-The research code that produces and consumes this data lives in the companion GitHub repository [`xiaomoguhz/R3-Bench`](https://github.com/xiaomoguhz/R3-Bench) (currently private; reach out for access if needed).
+```bibtex
+@article{r3bench2026,
+  title={Benchmarking and Evolving Reason-Reflect-Rectify for Reflective Visual Generation},
+  author={Anonymous},
+  journal={ICML (under review)},
+  year={2026}
+}
+```
 
 ## License
 
-Released under **CC-BY-4.0**. You are free to share and adapt, provided you give appropriate credit.
+Released under the **Apache 2.0** license. Source prompts adapted from T2I-CompBench, GenEval++, and GEdit-Bench under their respective licenses; see the code repository for full attribution.