schneiderkamplab/sapient-synth-tasksource-pragmeval-sarcasm
sapient-synth-tasksource-pragmeval-sarcasm Chat-template-ready synthetic anonymous replacement examples for one Sapient source excluded from the DFM5 data mix. Contents Format: gzip-compressed JSON Lines under data/train.jsonl.gz Schema: {"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]} Files: 1 Rows: 3753 Task: synthetic anonymous instruction replacement Generation Rows were generated with… See the full description on the dataset page: https://huggingface.co/datasets/schneiderkamplab/sapient-synth-tasksource-pragmeval-sarcasm.
sapient-synth-tasksource-pragmeval-sarcasm
Chat-template-ready synthetic anonymous replacement examples for one Sapient source excluded from the DFM5 data mix.
Contents
- Format: gzip-compressed JSON Lines under
data/train.jsonl.gz - Schema:
{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]} - Files:
1 - Rows:
3753 - Task: synthetic anonymous instruction replacement
Generation
Rows were generated with `google/gemma-4-31B-it` and accepted only after a `google/gemma-4-31B-it` judge checked task preservation, PII absence, low textual overlap, and usefulness.
Overlap precautions: prompts asked the model to create new synthetic examples that preserve the task type and skill, not to rewrite or copy the provenance row. Candidates were rejected when the judge found substantial wording overlap, unresolved PII, or failure to preserve the task. Only accepted rows are included.
License
Released as Apache-2.0. The datasets that inspired these recreations may have different licensing conditions. These examples are fully synthetic recreations with no or minimal wording overlap and were judged to be free of PII before inclusion.
Recreate
python recreate_dataset.py --validate
python recreate_dataset.py --from-synth /path/to/synth/tasksource_pragmeval_sarcasm.parquet --output /path/to/output-folder