GuangsTrip/surgbench_pipeline_dest_data

SurgBench file_pipeline smoke fixture

This is a small smoke/integration fixture for SurgBench file_pipeline. It is not a training set and not the complete SurgBench benchmark. It exercises per-file quotas, multi-file scheduling, cross-source image pools, resume behavior, and multi-model-group processing.

It contains 18 TSV files and 180 train samples: three sources × two shared tasks × three shards, with 10 unique rows per shard. Tasks: PhaseRecognition, StepRecognition. Sources: BernBypass_Laparoscopic, StrasBypass_Laparoscopic, PSI-AVA_Laparoscopic. Every source/task pair has 30 samples and 30 distinct decodable images.

Schema

UTF-8, tab-delimited TSV columns are:

image_path, question, A, B, C, D, gold_index, split, category, dataset, sample_id

gold_index is explicitly 0-based (0–3). image_path is relative to its TSV directory.

Local use

hf download GuangsTrip/surgbench_pipeline_dest_data --repo-type dataset --local-dir ./file-pipeline-smoke

# Recommended validation/planning pass first.
python -m file_pipeline --input-dir ./file-pipeline-smoke --dry-run

# Then use low concurrency and a small target ratio.
python -m file_pipeline --input-dir ./file-pipeline-smoke --max-workers 1 --target-ratio 0.1

If the local checkout exposes file_pipeline through a wrapper, keep this input directory and map the options to equivalent local flags.

Provenance, trimming, and access

Rows come from the authorized SurgBench source TSV collection. Only train files for the two selected tasks were considered. Original questions, option order, answers, and image bytes are retained; source base64 was decoded to JPEG/PNG/WebP. Test/validation/dev data, invalid rows, and duplicates were excluded. No full source TSV, model output, credentials, cache, citations, diagnostics, or private filesystem path is included.

The data owner/user explicitly authorized public upload on 2026-09-16. The repository is public. license: other is used because several source datasets are represented; downstream users must respect source-specific terms and should not infer broader relicensing of complete upstream datasets.

Distribution

TSV	Rows	Referenced images
BernBypass_Laparoscopic/PhaseRecognition_01.tsv	10	10
BernBypass_Laparoscopic/PhaseRecognition_02.tsv	10	10
BernBypass_Laparoscopic/PhaseRecognition_03.tsv	10	10
BernBypass_Laparoscopic/StepRecognition_01.tsv	10	10
BernBypass_Laparoscopic/StepRecognition_02.tsv	10	10
BernBypass_Laparoscopic/StepRecognition_03.tsv	10	10
StrasBypass_Laparoscopic/PhaseRecognition_01.tsv	10	10
StrasBypass_Laparoscopic/PhaseRecognition_02.tsv	10	10
StrasBypass_Laparoscopic/PhaseRecognition_03.tsv	10	10
StrasBypass_Laparoscopic/StepRecognition_01.tsv	10	10
StrasBypass_Laparoscopic/StepRecognition_02.tsv	10	10
StrasBypass_Laparoscopic/StepRecognition_03.tsv	10	10
PSI-AVA_Laparoscopic/PhaseRecognition_01.tsv	10	10
PSI-AVA_Laparoscopic/PhaseRecognition_02.tsv	10	10
PSI-AVA_Laparoscopic/PhaseRecognition_03.tsv	10	10
PSI-AVA_Laparoscopic/StepRecognition_01.tsv	10	10
PSI-AVA_Laparoscopic/StepRecognition_02.tsv	10	10
PSI-AVA_Laparoscopic/StepRecognition_03.tsv	10	10

See manifest.json for machine-readable counts, provenance, filtering rules, version, and generation time.