BraintrustDataDev/aid-evaluation-results

AID scene classification evaluation results

30,500 saved prediction records from five models evaluated on AID aerial scene classification. This release contains numeric predictions, reference-label identifiers, source references, hashes, prompts, protocols, and recomputed accuracy counts. It contains no AID image bytes, thumbnails, contact sheets, or transformed images.

The original images could not be cleared for redistribution. The upstream Hugging Face repositories declare an unknown license, and the original dataset consists of Google Earth imagery. Public availability of a mirror does not establish redistribution permission. See RIGHTS.md for the checked sources and release boundary. This repository does not grant a license to upstream imagery or annotations, or claim that the original benchmark is openly licensed. The metadata remains license: unknown; no blanket open license is asserted.

Included studies

Configuration	Predictions	Original images	Conditions
native_baseline	10,000	2,000	Historical short-prompt run
class_definitions	10,000	The same 2,000	Later run with class definitions
controlled_definitions	10,500	300 of those 2,000	Fresh control plus six transformations

Each study includes Luna, Jev, Clef, D1, and D1 Ultra, with one saved final prediction per case and model. Requested and served model identifiers are retained as recorded. These are historical experiment identifiers, not a claim of present public model availability. The five-model rows are correlated measurements, not 30,500 independent images.

The release covers the completed native baseline, class-definition follow-up, and definitions-matched robustness studies. Earlier duplicate reruns, pilot runs, the older robustness study without definitions, specialist checkpoints, paired-sharpness tasks, Wikimedia images, travel photographs, and the non-blind shared-error review are outside this release.

Load

from datasets import load_dataset

predictions = load_dataset(
    "BraintrustDataDev/aid-evaluation-results",
    "class_definitions",
    split="test",
)

Use either other configuration name to load that study. The test designation is inherited from a fixed community partition; it is not an independent train/test split of the prediction table. Keep every model, prompt condition, and image transformation for a source together when resampling or dividing these records.

Files and fields

• data/*.parquet: typed prediction tables used by the dataset loader. Matching data/*.jsonl files contain the same allowlisted records; raw provider responses and internal attachment records are omitted.
• references/test_images.jsonl: 2,000 image identifiers, upstream class references, byte hashes, dimensions, and pinned upstream paths. No images are downloaded by loading this release.
• prompts/*.txt: exact instructions used for each study.
• protocols/*.json: label order and frozen study protocols; local absolute paths are replaced with basenames. Preparation-time status fields describe the frozen plan; release_summary.json records the completed exports.
• metrics.json: independently recomputed counts and accuracy by study/model/condition. Invalid outputs count as incorrect.
• release_summary.json: exact release counts, source prediction/manifest checksums, and verification results for all 15 runs.
• SHA256SUMS.txt: checksums of the published release files, excluding itself.

Join records on original_id for the source image and on model for paired model comparisons. case_id identifies the actual original or transformed input. condition is original for native runs, and one of control, blur_1, blur_2, contrast_0.5, contrast_1.5, shift_left_60, or shift_right_60 for robustness.

reference_label records the upstream category, not a new adjudication. predicted_label, status, valid, correct, and the 30-label probabilities describe the parsed result. Invalid rows retain their status; probabilities are null. input_sha256 hashes exact submitted image bytes; original_sha256 hashes the original source bytes. prompt_sha256 hashes exact instruction text. dataset_version identifies the frozen Braintrust dataset snapshot.

latency_ms and attempts are historical runner measurements. Token counts and costs may be null when unavailable. provider_reported_cost_usd is read from numeric provider usage.cost; recorded_cost_usd, token_estimated_cost_usd, and cost_source retain the saved accounting fields. legacy_estimated_api_cost_usd preserves the older field without relabeling it as a billed amount. Do not sum multiple cost columns. Costs are historical and may omit unreported failed-attempt charges; they are not current price quotations or invoices. Missing fields are not zeros.

Results and interpretation

Model	Native baseline accuracy	Class-definition accuracy
Luna	74.30%	85.55%
Jev	78.95%	86.15%
Clef	78.05%	83.85%
D1	77.25%	84.30%
D1 Ultra	78.95%	85.85%

The class-definition study was designed after inspecting baseline errors. Some definitions are working interpretations rather than official AID rules. Comparison with a historical baseline cannot isolate prompt effects from intervening provider or model changes. There is one trial per case; these records do not establish run-to-run variability or an unbiased confirmatory benchmark ranking.

The robustness study uses ten images per class, selected by the lowest SHA256 of aid-controlled-20261022-v1 concatenated with the filename. All seven conditions use the same definitions prompt. Controls and transformed images are lossless 600×600 RGB PNGs. Blur uses Pillow GaussianBlur radii 1 and 2; contrast uses ImageEnhance.Contrast factors 0.5 and 1.5; horizontal shifts use 60-pixel reflected padding and remove 60 original boundary pixels. Exact recipes are documented in the frozen protocol. The original source bytes and resulting image hashes are separate because PNG re-encoding changes bytes. Pixel-equivalent re-encoding with another library version need not produce identical file hashes.

Compare robustness conditions against the fresh control rows for the same source and model. Full-test accuracy and historical native predictions are not valid controls for these transformations. The balanced 300-image subset has a different class distribution from the full test split. For statistical analysis, cluster by original image and preserve paired conditions; confidence intervals from these rows alone do not measure repeated-call variation. Reference categories can be ambiguous and are not corrected by model consensus.

Provenance

Original benchmark: AID, Xia et al., 10,000 images and 30 scene categories. Source image references use blanchon/AID, revision d9b5531544523f29d365a77a007b2ace3cea6882. The 6,000/2,000/2,000 community train/validation/test partition comes from isaaccorley/aid, revision e6767964e40f567a18eac6e6ca5fce397e4ce411. This is not the original paper's repeated random-split benchmark protocol.

Evaluation project: aid-single-label-comparison. Native run versions: 1000198073523310044; controlled run version: 1000198078684663546. All prediction IDs, input hashes, row counts, valid counts, and aggregate accuracies were checked against the frozen local manifests and saved summaries before publication. This verification does not independently validate the upstream category judgments.

Please cite the original dataset when discussing results:

@article{xia2017aid,
  title={AID: A Benchmark Data Set for Performance Evaluation of Aerial Scene Classification},
  author={Xia, Gui-Song and Hu, Jingwen and Hu, Fan and Shi, Baoguang and Bai, Xiang and Zhong, Yanfei and Zhang, Liangpei and Lu, Xiaoqiang},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  volume={55}, number={7}, pages={3965--3981}, year={2017}
}