Most datasets supply inputs and a single target label; this corpus supplies typed decision rows that preserve the original state, question, ordered options and numeric decision targets so models can return calibrated probabilities and multiple decision heads.
What Sets It Apart
- Explicit typed decisions: each row encodes Choice (probability over candidates), Noul (yes/no probability) or Score (ordinal/numeric), enabling direct probability estimation and calibration workflows.
- Multi-config frozen releases: twelve frozen configs and dozens of splits (train/calibration/validation/test/OOD) cover controlled domains such as customer-support routing, simple games (Snake, tic-tac-toe), simulated browser/drone controls, citation verification, entity-alignment and extraction tasks. Example counts: release-v2-redistributable (79,116 train; 113,568 total), browser-drone expansion (108,624 train; 160,797 total); the original frozen release used to train 2B/9B runs had 80,816 train rows and 115,821 rows across splits.
- Reproducibility and provenance: raw JSONL.gz, Parquet exports, original manifests, audit artifacts and a REPRODUCTION.md let you reconstruct exact mixtures; generated records are CC0-1.0 and source code is MIT.
- Serving and evaluation focus: records are designed for decision-model training, calibration and inspection (the Open-Jev decision interface reads one forward pass per question and returns calibrated option probabilities). Guidance is explicit about what fields can be used as model input (use state, question, kind, options; do not expose metadata/targets to models).
Who it's for and trade-offs
Great fit if you are training, calibrating or auditing models that must output explicit decision probabilities (e.g., decision heads, rerankers, agent action selectors) or benchmarking calibration and multi-head decision behavior. Also useful for building and testing deterministic candidate-extraction + typed selection pipelines. Look elsewhere if you need large-scale natural web scrapes, private mailbox contents, multimedia raw videos, or production-grade flight control data — the tasks are controlled/synthetic and several control corpora have not been used for model training. Some configs overlap and certain source rows (e.g., Wikispeedia) are excluded from public projections, so reproduce exact frozen mixtures if you need byte-for-byte parity.
