QWE

Speed (SKT AI LABS) SFT training dataset

Instruction-tuning dataset with chain-of-thought, merged and deduplicated from multiple public reasoning datasets. Every row is JSON with:

• instruction - the user/task prompt
• thought - reasoning wrapped in `` tags (always present)
• response - the final answer
• has_thinking - always true
• source_dataset - provenance of the row

The data is sharded: data/train-00000.jsonl.gz ... data/train-NNNNN.jsonl.gz in row order. Sources include OpenThoughts3-1.2M, Mixture-of-Thoughts, Infinity-Instruct and Nemotron-Post-Training-Dataset (v1 and v2). Deduplication is exact full-line (md5), applied globally across all sources.

Notes

• All model-identity references were normalized to "Speed by SKT AI LABS Private Limited".
• A small fraction of rows (mostly from one legacy source) carry a placeholder thought; filter them with --skip-placeholder when building training data.
• nvidia/Nemotron-Post-Training-Dataset-v1 is only partially included: the ingestion process was interrupted partway through the v1 math shards.