Self-labeled training data used to train PeetPedro/kompress-v4.
1802 pairs from kompress_train_split.jsonl, references replaced by kompress-v3+override compressed output. mk_in_ref = 0.823 (vs ~0.5 for original Q&A labels).
Each row: {"text": str, "reference": str, "role": str, "source": str, "topic": str}
The self-labeling step that produced this dataset is the key insight of the ultrawhale loop: use the previous model's compressed output as the training… See the full description on the dataset page:
https://huggingface.co/datasets/PeetPedro/kompress-v4-traindata.
