Jaqen0-0/SafeLoop-Training-Data

SafeLoop v56 Training Data

This dataset contains the public training data used for the SafeLoop v56 predictor-head release. It pairs three JSONL training files with a de-duplicated set of rollout image frames stored as tar shards.

Code: https://github.com/Loule0-0/SafeLoop

Weights: https://huggingface.co/Jaqen0-0/SafeLoop

Contents

• data/hardcase_v9_with20260706_controller_ablation.jsonl: object-recall stage data.
• data/merged_v23_large_clean.jsonl: success-balanced stage data.
• data/merged_v30_newstuck_oversample4.jsonl: final low-drift stuck/object stage data.
• shards/images_*.tar: image frames referenced by the JSONL files, preserving their relative paths.
• manifest.json: counts, checksums, task suite, and shard metadata.

Use

Download this dataset, extract the shards with the repository helper, and set the trainer data root to the extracted directory.

hf download Jaqen0-0/SafeLoop-Training-Data --repo-type dataset --local-dir $SAFELOOP_DATA
python scripts/materialize_hf_training_data.py --dataset-dir $SAFELOOP_DATA --out $SAFELOOP_DATA/materialized
python scripts/run_release_predictor_training.py --recipe v30_lowdrift_success --data-root $SAFELOOP_DATA/materialized --model-dir $QWEN_MODEL --weights-dir $SAFELOOP_WEIGHTS --output-root $SAFELOOP_OUTPUT

The base VLA policy and Qwen2.5-VL backbone are not redistributed here.