Recommended for AI training. This version replaces the template-based builds and is fully LLM-generated for true independence.
Previous builds (30K and 10K-template) were template-based (gen.py + 178 situations):
13,156 distinct messages from 74,526 total → 82.3% duplicate messages, max x109 on narrative blocks (Mesa seis pegada a la mesa siete...)
Coherence failures from… See the full description on the dataset page:
https://huggingface.co/datasets/itsZyn/ConvES30K.
