# Caption catalogue — S9 data bundle

`caption_catalogue.jsonl` is a **1,000-row synthetic caption catalogue**: 800 rows in the `train` split and 200 in `validation`. It is 464 KB.

## Where it comes from

Course-generated, not downloaded. The deterministic generator is the prototype's `build_data.py` (seed 42) under `prototypes/ai-inference-finetuning/`; the same generator is embedded in `qlora_lab.py`. Subjects are objects, campus spaces and 3D assets; there are **no people, faces, brands or client footage**.

## Licence and reuse

Course-owned synthetic material. Reuse within the course and for teaching; do not present it as real client data. The text may be regenerated from the generator, so it is not a fixed external dataset.

## Fields

Each line has `id`, `split` (`train` / `validation`), `metadata` (asset type, subject, setting, mood, palette, platform, audience, format, licence), `title` and `caption`.

## How to get it

- **Local / MLX:** it is already at `data/caption_catalogue.jsonl` in this week's folder.
- **Kaggle:** create a **private dataset** containing this file, then attach it under *Add Data*. The notebook searches `/kaggle/input` for `caption_catalogue.jsonl`, so the dataset name and mount layout do not matter; set `COURSE_DATA` to the folder holding the file if you want to be explicit. Files written to `/kaggle/working` do not survive the session.
- **Faculty RTX / RunPod:** copy the same file to the working directory.

No runtime download and no network fetch is required.
