Grounding-EvalData

Grounding-EvalData is distributed by academic researchers solely for non-commercial scholarly research, benchmarking, and reproducibility. It provides the complete evaluation inputs for the 34 benchmark entries in GroundingPI. GroundAnything uses the documented 30-entry subset.

All materials remain subject to their original upstream licenses and restrictions. This collection does not grant a unified license over third-party data. Read LICENSE and THIRD_PARTY_NOTICES.md. The release is provided as is, without warranties or liability to the maximum extent permitted by applicable law.

Contents

The repository follows the Rex-Omni-EvalData organization: source archives at the root and JSONL annotations under _annotations/.

Grounding-EvalData/
├── README.md, LICENSE, LICENSE-CODE, THIRD_PARTY_NOTICES.md
├── licenses/                       # Upstream terms and provenance
├── _annotations/
│   ├── box_eval/                   # 13 JSONL files
│   ├── point_eval/                 # 7 JSONL files, including mask RLE
│   └── visual_prompt_eval/         # 4 JSONL files
├── coco.tar.gz, dense200.tar.gz, ...# 11 unchanged Rex image archives
├── refcoco.tar.gz, refcocoplus.tar.gz, refcocog.tar.gz
├── refspatial_bench.tar.gz, robospatial_home.tar.gz
├── screenspot_pro.tar.gz, screenspot_v2.tar.gz, osworld_g.tar.gz
├── support/OCR/                    # ICDAR2015 ignore-region labels
├── configs/suites.json             # Exact 34 / 30 task selections
├── scripts/                       # Verify, unpack and configure
├── BENCHMARKS.csv, BENCHMARKS.json, DATA_SOURCES.json
├── EVALUATION_FILES.jsonl, package_manifest.json, SHA256SUMS
└── validation/                    # Readability and identity checks

The 34 entries cover 20 underlying dataset families; they are not 34 independent image corpora. RefCOCO family averages and RefSpatial averages are not extra tasks. The 8 additional registry recipes (*_Labelless and *_Boxonly) are evaluation variants sharing these inputs. Training data and unrelated source splits are outside this evaluation release. Every record and asset consumed by the released evaluator is included: no additional subsampling or smoke limit is applied. Some Rex selections differ from the original benchmark sizes (for example, COCO box evaluation uses 4,952 records and ICDAR2015 uses 496); the supplied source annotations and counts are preserved.

Download and connect to the evaluation code

Install the Hugging Face download utility and download the repository:

python -m pip install -U huggingface_hub hf_xet
hf download Skywalker0410/Grounding-EvalData \
  --repo-type dataset --local-dir ./Grounding-EvalData

Keep the download directory; its loose annotations remain part of the evaluation inputs. Use a separate extraction directory and an existing checkout of either released code repository:

python Grounding-EvalData/scripts/prepare_data.py \
  --data-root /absolute/path/to/grounding-eval-data \
  --code-root /absolute/path/to/GroundingPI

For GroundAnything, replace the code root. Both can be configured in the same invocation by repeating --code-root. The standard-library helper verifies SHA256SUMS, extracts all 19 archives, checks data paths and writes configs/datasets.local.yaml. It preserves existing source/config files and refuses to replace different data. Allow approximately 18 GB for the downloaded repository and 19 GB for extracted data; retain extra space for models and evaluation outputs. Python 3.9+ is sufficient for unpacking; the release evaluator uses its own documented environment. Parquet validation was performed with PyArrow 21.0.0; 19.0.0 is incompatible with some supplied RefSpatial/RoboSpatial parquet metadata.

If an existing evaluation environment has PyArrow 19 or raises Repetition level histogram size mismatch, use the tested reader version without rewriting the data:

.venv-eval/bin/python -m pip install "pyarrow==21.0.0"

Create a full-suite evaluation recipe from the corresponding official template, using the evaluator environment (which includes PyYAML):

# Run from the GroundingPI checkout after its documented environment setup.
.venv-eval/bin/python /absolute/path/to/Grounding-EvalData/scripts/create_eval_config.py \
  --code-root . --template configs/eval/gam.yaml \
  --data-root /absolute/path/to/grounding-eval-data \
  --suite groundingpi34 --output configs/eval/full34.local.yaml

.venv-eval/bin/python scripts/evaluate.py configs/eval/full34.local.yaml --dry-run

For GroundAnything use --suite groundanything30 and its service-matching template: configs/eval/dlm.yaml, configs/eval/dlm_speculative.yaml, or configs/eval/vlm.yaml. These correspond to denoising, self-speculative decoding and the separate VLM checkpoint. The helper retains model/service settings from the selected template and assigns a fresh run ID; inspect those settings against your running service. All released checkpoints use mode: GAM. Full evaluation uses limit: null.

After the dry run shows no missing inputs, start the matching model service according to the code repository and execute the same evaluator command without --dry-run. A missing model directory is a separate model installation issue, not a missing dataset. No model weights or inference outputs are included here.

The tested path mapping keeps Rex images under <data-root>/images/<source>/..., source parquet/GUI directories under <data-root>/<dataset>/..., and points annotation keys directly at this repository's _annotations/. Raw JSONL dynamic category dictionaries are preserved; use the supplied evaluator instead of combining all files with a generic Arrow JSON loader. The HF viewer is disabled for this heterogeneous collection.

Benchmark inventory

raw_rows counts stored annotation records; evaluation_rows counts evaluator-expanded requests. RefCOCO expressions and visual-prompt categories can produce multiple requests per record. Values describe the bundled inputs, not reported model scores.

Task	Raw rows	Evaluation rows	GroundAnything 30
gam_humanref	5,000	5,000	—
gam_refcocog_val	4,889	4,889	yes
gam_refcocog_test	9,577	9,577	yes
gam_refcoco	8,811	25,080	yes
gam_refcocog	7,573	14,432	yes
gam_refcocoplus	3,805	10,758	yes
gam_coco	4,952	4,952	yes
gam_lvis	19,626	19,626	yes
gam_dense200	200	200	yes
gam_visdrone	1,610	1,610	yes
gam_refspatial_location	100	100	yes
gam_refspatial_placement	100	100	yes
gam_refspatial_unseen	77	77	yes
gam_robospatial_context	122	122	yes
gam_rex_point_humanref	4,997	4,997	—
gam_rex_point_refcocog_test	9,572	9,572	yes
gam_rex_point_refcocog_val	4,887	4,887	yes
gam_rex_point_coco	4,950	4,950	yes
gam_rex_point_lvis	19,615	19,615	yes
gam_rex_point_dense200	200	200	yes
gam_rex_point_visdrone	1,610	1,610	yes
gam_screenspot_pro	1,581	1,581	yes
gam_screenspot_v2	1,272	1,272	yes
gam_osworld_g	564	564	yes
gam_hiertext	1,723	1,723	yes
gam_icdar2015	496	496	yes
gam_totaltext	300	300	yes
gam_sroie	360	360	yes
gam_doclaynet	6,480	6,480	yes
gam_m6doc	2,724	2,724	yes
gam_fsc147	1,190	1,190	yes
gam_visual_coco	4,952	14,631	—
gam_visual_dense200	200	200	yes
gam_visual_lvis	19,626	70,139	—

Verification and provenance

All 34 selected tasks have complete input paths. Validation fully decoded pixels with Pillow Image.load() for 46,008 external image/mask files, 20,588 embedded images and 399 embedded masks in 13 parquet files, with no errors and without truncated-image recovery. All 277 referenced RefSpatial external masks match the embedded masks. One additional upstream mask is retained. Full extraction and source-byte checks are recorded in validation/; this is data validation, not model inference or metric reproduction.

The Rex files are pinned to revision daff862ba483f56ccaba9af5a12e118a9372ede8. The release path contract was examined against GroundAnything commit 5534c0043564632120c663e4b64b7c700e624e6e and GroundingPI commit a7804908bce5fb6228fb9ae1c729b7461240d1ce. The official 30-task and 34-task evaluator dry runs both resolved all 39 path keys without missing dataset inputs; only model weight directories were absent in the validation environment. Local checkpoint paths, training data, caches, logs, credentials and model predictions generated during local evaluation are not bundled. Some original upstream JSONLs contain a legacy predict field; those source records are preserved unchanged and are not new predictions from this release.

Please cite the relevant code/model publications and the original dataset authors listed in THIRD_PARTY_NOTICES.md. This aggregation does not replace attribution to the original benchmarks.