puttatidam/sib200-lao-clustering
sib200-lao-clustering Deduplicated copy of kornwtp/sib200-lao-clustering, part of the SEA-BED data-quality work. Source dataset: kornwtp/sib200-lao-clustering Deduplicated on: 2026-09-04 Task type: clustering Splits: test, validation What changed A text appearing under more than one gold cluster is collapsed to ONE row carrying the competing cluster ids in a new conflict_label column (NULL elsewhere). labels on that row holds the first-seen value as a… See the full description on the dataset page: https://huggingface.co/datasets/puttatidam/sib200-lao-clustering.
sib200-lao-clustering
Deduplicated copy of `kornwtp/sib200-lao-clustering`, part of the SEA-BED data-quality work.
- Source dataset:
kornwtp/sib200-lao-clustering - Deduplicated on: 2026-09-04
- Task type: clustering
- Splits: test, validation
What changed
A text appearing under more than one gold cluster is collapsed to ONE row carrying the competing cluster ids in a new conflict_label column (NULL elsewhere). labels on that row holds the first-seen value as a placeholder -- it is not a decision; a downstream human/LLM pass picks from conflict_label and writes the verdict back into labels. Collapsing means one row and one closed candidate set per decision, instead of the same choice repeated once per competing label with nowhere to record an answer. Duplicate texts that AGREE on their cluster are left in place: redundant, not wrong.
Row counts and the exact policy used are recorded in clustering_dedup_summary.csv in the analysis repo. This dataset is not identical to the published SEA-BED data -- scores computed on it are not directly comparable to published SEA-BED numbers.