AgentEndeavour/omni-2stage-news
Omni 2-Stage News Trading Dataset Two-stage training data for a pre-open equity direction model. Stage 1 — Pure-news impact (PPO) stage1_impact/news_reward_map.jsonl (68880 rows). One news item (title + licensed summary) that may move many names. Label fields per row: dir64 / r64_abn: realised ~2.6-day abnormal direction & magnitude of the focal name, block_<id>_dir: realised direction of every affected sector block, big_up64/big_dn64: excursion flags (magnitude… See the full description on the dataset page: https://huggingface.co/datasets/AgentEndeavour/omni-2stage-news.
Omni 2-Stage News Trading Dataset
Two-stage training data for a pre-open equity direction model.
Stage 1 — Pure-news impact (PPO)
stage1_impact/news_reward_map.jsonl (68880 rows). One news item (title + licensed summary) that may move many names. Label fields per row:
dir64/r64_abn: realised ~2.6-day abnormal direction & magnitude of the focal name,block_<id>_dir: realised direction of every affected sector block,big_up64/big_dn64: excursion flags (magnitude head),cs_realized: surface-read vs realised-move mismatch (trap / counter-signal).
Reward = realised move (outcome) + critic consistency against the licensed "why" summary.
Stage 2 — Full-flow decision (GRPO)
stage2_decide/{train,val}.jsonl (168/54 decision timepoints). Each row carries the ENTIRE past-24h news flow (incl. pre-market mover lists), the visible trading state of every candidate, and the realised next-24h per-name label (labels.<sym>.direction / r24_abn / big_*). GRPO has no critic; reward is the realised move only.
All inputs are strictly point-in-time; nothing forward-looking enters a prompt. License: research use.