QRaco Physical AI Recorder dataset

Primary data: records_part_*.jsonl.gz
Storage profile: compact primary sensors, source Float32 precision preserved.
Derived Android streams are omitted because gravity and linear acceleration can be reconstructed offline.
Protocol and device metadata: metadata.json
Field contract: schema.json
Compact integrity chain: blocks.jsonl
File checksums: integrity.json (created when ZIP is exported)
Auto-detected rare events are heuristic candidates, not ground truth.
Confirm important events with manual markers and document the collection protocol.

Parquet conversion on a PC:
  1. Install Python and run: pip install pyarrow
  2. Put convert_to_parquet.py beside the JSONL.GZ parts.
  3. Run: python convert_to_parquet.py

The converter is streaming and writes records.parquet with ZSTD compression.