The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
Author-scale 590-row validation data rescue
Actual multimodal parquet payload: 590 rows / 69151663 bytes, transported in two <=64MiB parts. Includes original questions, response pairs, image bytes, published preference/safety labels and prepared source IDs/image hashes/lengths. All590 logical rows survived the minimum credential/contact and image-GPS checks; zero rows excluded. Images remain original upstream bytes. No model/optimizer states included.
Upstream attribution: saferlhf-v/BeaverTails-V, fixed revision ee19041205c720c0faea575de563de8a6a8f9094; CC-BY-NC-4.0 (noncommercial, attribution and applicable adaptation terms). Source card: https://huggingface.co/datasets/saferlhf-v/BeaverTails-V/blob/ee19041205c720c0faea575de563de8a6a8f9094/README.md . This safety research dataset contains potentially disturbing/offensive prompts and answers, not endorsement or certified safe model behavior. Existing official evaluation has been reused historically and is not a fresh blind test or human-gold certification.
This reserialized reviewed parquet is byte-different from the historical69062018-byte source while retaining all590 logical rows. The publication is not a claim that all15.21GB of training/RL/PTX data are backed up. ROW_MAPPING and RECOVERY_MANIFEST state exact coverage, provenance and scan limits.
Restore
hf download wangzyuan/SafeRLHFV-Author-Scale-Validation-Rescue --repo-type dataset --revision COMMIT --local-dir ./validation-rescue
python ./validation-rescue/restore_research_data.py ./validation-rescue
The restore program verifies each part and the reconstructed69151663-byte SHA256 before publishing restored/validation-reviewed-subset.parquet. It refuses to overwrite existing outputs. Read with PyArrow/Pandas; no training or remote dataset code is run by reconstruction. The fixed commit comes from the publication receipt. Checks cover all text cells for credential/contact patterns and all image GPS EXIF, not a universal automated image privacy certification.
- Downloads last month
- 41