LoRA Training

Data

1. Flatten log (ecosystemruntime_flatten): offers that were shown, with accepted as the label.
2. Offer matrix: every offer that may be presented gets an item vector at export time, including cold offers never shown in the log.
3. Templates: user and item text templates filled from selected columns (customer, channel, segment, offer attributes).

Steps in Workbench

1. Create a config on LoRA Recommender.
2. Set rank, alpha, and templates in the UI. Use rank 0 when you want baseline encoder vectors without the trained residual.
3. Train — runs the configured training job in Workbench (notebooks sidecar when enabled).
4. Concept test — rank offers for one customer offline.
5. Export embeddings — write logging.lora_user_embeddings and logging.lora_item_embeddings for the run_id.

Training reads accept/reject signal from the flatten store; export must cover all matrix offers you intend to score live.