| # Schema |
|
|
| Each JSONL record in `metadata/` contains the following fields: |
|
|
| - `id`: unique sample identifier. |
| - `split`: dataset split (`train`, `val`, `test`). |
| - `domain`: one of `finance`, `medical`, or `iiot`. |
| - `video`: relative path to the MP4 file. |
| - `audio`: relative path to the WAV file. |
| - `frame_preview`: relative path to a representative PNG frame. |
| - `duration_sec`: clip duration in seconds. |
| - `fps`: video frame rate. |
| - `sample_rate`: audio sampling rate. |
| - `task_type`: benchmark task family. |
| - `question`: generated question. |
| - `ground_truth`: deterministic answer. |
| - `has_conflict`: whether a counterfactual cross-modal conflict was injected. |
| - `source_mode`: substrate source (`synthetic` or `yfinance` for finance in the full generator). |
| - `seed`: random seed used by procedural generation. |
| - `metadata_file`: relative pointer to the split metadata file. |
| - `qa_file`: relative pointer to the split QA file. |
| - `algorithmic_metadata`: domain-specific gradients and trigger values used by the state machine. |
|
|
| The `qa/` JSONL files mirror the QA-specific subset of this information. |
|
|
