cnfeichen/t2reranking-4level-judgments

T2Reranking 4-level relevance judgments (teacher-graded)

17,390 four-level relevance judgments over the C-MTEB T2Reranking dev hard-negative pools: for every (query, passage) pair, a probability distribution over irrelevant / related / partial / answers — the 4-level rubric from awesome-jev-prompts's relevance-score patterns, graded by deepseek-v4-flash in pooled per-query prompts (all 10 passages per call, temperature 0.2, thinking disabled). Graded 2026-10-21; full generation script with cost discipline (dry-run, circuit breakers, resume) lives in the minimind-jev repo.

No query or passage text is included. Records are keyed (qid, role, idx) and join 1:1 with the splits produced by minimind-jev's scripts/prepare_data.py (seed 42) — run it to regenerate the text side, pulling T2Reranking yourself under its own terms.

Quality

Measured against the dataset's own binary labels:

• argmax agreement 81.8% — the rubric is finer than binary gold, so part of the gap is granularity, not error;
• expected-score separation: positives mean 0.557 vs negatives 0.322, AUC ≈ 0.716 — the same band as hosted Jev on the same kind of pools (0.714–0.718).

One honest caveat

Distilling these labels into a small head over frozen bge-small-zh features reduced ranking quality relative to plain binary training (nDCG@5 0.488 → 0.428–0.430, three configs, paired CIs excluding zero) — full analysis. The judgments are most useful as an independent signal — threshold studies, rubric-shape analysis, weak supervision where no gold labels exist — not as a drop-in upgrade for gold-ranking objectives.

Format

One JSON object per line:

{"qid": "q00123", "role": "positive", "idx": -1, "hard": 1,
 "dist": {"irrelevant": 0.02, "related": 0.08, "partial": 0.30, "answers": 0.60},
 "teacher": "deepseek-v4-flash"}

role/idx identify the passage within its query pool (positive / negative 0..8); hard is the dataset's own binary label; each dist sums to 1.

Provenance & license

• Graded 2026-10-21 (off-peak), 1,739 pooled calls, zero parse failures; 5.98M input + 0.60M output tokens, ~¥8 total.
• The judgments are released under CC-BY-4.0. The underlying T2Reranking text is not redistributed.
• Part of the minimind-jev project (train-your-own tiny Jev); companion checkpoint: cnfeichen/minimind-jev-v0.1.