遇见数据集

PACTS Corpus V6 External-Holdout Evaluation: Release-Safe Replication Package

收藏
Zenodo2026-07-06 更新2026-08-02 收录
官方服务:

资源简介:

Release-safe replication package for the locked V6 external-holdout evaluation of the PACTS Corpus project. The archive contains 34 evaluation-eligible record-level inputs, human-adjudicated final labels, adjudication-provenance limits, source linkage, 11 normalized FDA warning-letter source texts, Arm A and Arm B per-record scoring outputs, fixed-threshold evaluation metrics, paired-bootstrap uncertainty results, a portable reconciliation checker, an independent package validator, and SHA-256 integrity manifests. The V6 pre-lock population contained 35 records. V6-DRAFT-036 is retained in the exclusion log but excluded from the evaluation denominator because it lacked a retrievable normalized source and review excerpt. The eligible evaluation population contains 8 Yes and 26 No labels. Evaluation policy: fixed threshold 0.0; mean signed-margin seed aggregation; no post-V6 threshold tuning; no post-V6 model selection. Bootstrap analysis used 10,000 paired stratified replicates with seed 42. This release excludes raw HTML captures, discovery workbooks, working drafts, training datasets, model checkpoints, tokenizers, and private workstation paths.

提供机构:
Zenodo
创建时间:
2026-07-06
二维码
社区交流群
二维码
科研交流群
商业服务