The training data behind K2 Horizon, released by IFM under CC-BY-4.0 from August 24, 2026: TxT360-v2, the successor to TxT360 with web-high, medium, and QA subsets, plus Code-Reasoning (seven subsets), Math-Reasoning (five), SFT-Reasoning, and Pretrain-Behaviors. A PR-based agentic-code synthesis pipeline (PRism-synthesis) followed on GitHub in September. Filed late; the K2 Horizon entry had listed its training data as forthcoming.

Dataset

datasetopen-source

Related