# Reconstructing the winning training mix

The winners were trained on `mix_all.jsonl`, assembled on the training box and not preserved.
It is reconstructible from two files in this repo:

    val700_train.jsonl        7,988 rows   - from the released validation videos (53.8% interrupt)
    egoconv_dense_train.jsonl 13,730 rows  - the ambient-agent synthetic corpus (46.5% interrupt)

Mix = val700 rows twice, concatenated with the full egoconv set, shuffled:

    cat val700_train.jsonl val700_train.jsonl egoconv_dense_train.jsonl | shuf > mix_all.jsonl

Schema: `images`, `query`, `history`, `target`, `label`, `video`, `domain`, `chunk`, `system`.

## Frames

`images` holds paths **relative** to your own frame directory, as `<video_stem>/iv<NNN>/f<MM>.jpg`.
The frames are not published here (~57 GB of JPEGs); generate them from the source videos:

    python src/convert_train.py --rows egoconv_dense_rows.jsonl --out chunks.jsonl \
      --videos <videos> --frames_root <frames> --max_frames 32

That writes 16 frames per interval into `<frames>/<video_stem>/iv<NNN>/f<MM>.jpg`, which is exactly
the layout these rows expect. Then train with `--frames_root <frames>`.

## Training

Use `train_verbalizer.py` (NOT `train_proactive.py`, which is degenerate):

    python train_verbalizer.py --data mix_all.jsonl --frames_root <frames> \
      --out <adapter> --epochs 1 --rank 32 --alpha 64

Code: https://github.com/ambient-intelligence-labs (see the EgoProactive solution repo).