Datasets:

TypeSafeAI
/

Open-Jev

11,600 typed rows, 8,800 request cases, 400 queries and 200 synthetic system families. Original generated content and labels are CC0-1.0. Source code is MIT. This is a finite original relevance task, not TREC or scraped search data.

Each query asks for the current retry ceiling and backoff delay for one fictional system/profile. Grade 3 supplies both values; grade 2 supplies one; grade 1 is about the requested system but has the wrong profile, withdrawn guidance or no requested value; grade 0 concerns another system. A quoted query does not change the passage's subject. Two counterfactual queries share each family's eight passages.

Six request forms are supplied: pointwise Noul/Score, opposite-order pairwise Choice, local setwise Choice, listwise Choice and listwise Score. Each query has 22 request cases and 29 typed rows. One listwise Score request has eight questions. Score targets are four-level hard grades. Noul is positive only for a complete grade-3 answer. Choice targets are uniform over equally best passages; they do not supervise a full ranking or represent measured model uncertainty.

Split	Typed rows
train	7,366
calibration	464
validation	464
test	986
ood	2,320

All query variants, passages and methods from a system family stay together. OOD reserves brief wording and different profile names for 40 whole families. Family index modulo 10 values 8–9 designate OOD; other families use stable group hashing. Expansion preserves every frozen pilot family assignment. No pilot test/OOD family enters train, calibration or validation. Rows within a family are correlated and must not be presented as independent real-world examples.

Use only state, question, kind and options as model inputs. Targets, metadata, queries' reference relevance and auxiliary cases are privileged audit data. Each Parquet record_json and the decompressed raw JSONL preserve the original row bytes and order exactly, using the root card's existing column schema.

No training on this corpus is claimed. A separate frozen pilot received real Jev 1.13.0 outputs on six held-out queries (132 requests / 174 typed answers). That small pilot is described in the linked source documentation; it is not evaluation of all 11,600 rows, a TREC reproduction, or evidence of newly trained Open-Jev capability. Generation manifests' inference flags record generation-time operations, not the absence of subsequent pilot evaluation. The frozen release-v2 mixture and published model training data remain unchanged.

The community interface source supplies task shapes. Our prompts, passages and labels are independently authored. No third-party examples, TREC passages/qrels or provider responses are included.

From that pinned source checkout, use a fresh output directory:

python3 -m jev.ir_data --output-dir data/ir-control-v1-rebuilt --groups 200 --ood-groups 40 --seed 42
python3 reports/ir-control-v1/verify.py --data data/ir-control-v1-rebuilt --output ir-audit.json

External TREC evaluation remains pending. The local holdout loader isolates external candidates/qrels from the training-data tree and uses linear-gain trec_eval ndcg_cut semantics; no external benchmark result is bundled here.