Datasetpublic

schneiderkamplab/dfm10-mimir-drop-reasoning-sft

dfm10-mimir-drop-reasoning-sft Grounded discrete reading-comprehension examples with executable arithmetic supervision. Contents Format: gzip-compressed JSON Lines under data/train-*.jsonl.gz Schema: chat messages, optional condition and tools, plus provenance Shards: 1 Rows: 188,781 Category: Discrete reading comprehension Upstream material Novel openly grounded passages and generated questions Processing Gemma 4 31B generated… See the full description on the dataset page: https://huggingface.co/datasets/schneiderkamplab/dfm10-mimir-drop-reasoning-sft.

sourceHugging Faceotherupdated 27d agoView on Hugging Face

0likes158downloads

Dataset Card

dfm10-mimir-drop-reasoning-sft

Grounded discrete reading-comprehension examples with executable arithmetic supervision.

—Format: gzip-compressed JSON Lines under data/train-*.jsonl.gz
—Schema: chat messages, optional condition and tools, plus provenance
—Shards: 1
—Rows: 188,781
—Category: Discrete reading comprehension

Upstream material

—Novel openly grounded passages and generated questions

Processing

Gemma 4 31B generated candidates; operand grounding and arithmetic execution checks, independent five-dimension audit, exact decontamination, deduplication, and a 4,096-token cap were applied.

Selection policy: all rows in the packaged source artifact.

Every packaged row is taken from the accepted source tree identified in the package manifest. Tokenized arrays and epoch sampling indices are not included; export staging alone does not imply inclusion in a sampled training union.

License and release review

This package does not replace or broaden the licenses of its upstream materials. Review the dataset card, preserve upstream notices and attribution, and record the release decision before upload. Preserve row-level source provenance and upstream license metadata.

Validate

bash

python recreate_dataset.py