ceselder/

loracle-training-data-v2

Training data for the LoRA Oracle (loracle) project — a model that reads LoRA weight deltas and describes what behavioral changes the LoRA induces, without running the model.

	Dataset Contents




	
	rollouts_all.parquet (633K rows, 9980 prompts)

Behavioral rollouts for training LoRAs. Each prompt has ~64 conversation examples (32 trigger + 32 normal).

prompt_id: unique identifier linking to prompts_10k.parquet user_message: realistic user… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-training-data-v2.