遇见数据集

Replication Package for "AnchorGrad: Verifiable and Traceable LLM-Based Automated Assessment via Rubric-Front-Loaded Contracts"

收藏
Zenodo2026-11-01 更新2026-11-05 收录
官方服务:

资源简介:

This is the replication package for the IEEE TKDE submission "AnchorGrad: Verifiable and Traceable LLM-Based Automated Assessment via Rubric-Front-Loaded Contracts" (Yang Wei, Chunlin Xu, Xing Tian). AnchorGrad front-loads the marking scheme: a rubric anchor — scoring points with weights, a reference answer, a numeric self-check expression, and accepted equivalent forms — is emitted in the same LLM generation call as the test item and persisted in the same record, so that item decidability, score traceability, and failure auditability can be verified rather than assumed. The package contains: 1. Manuscript/ — the LaTeX source of the submission (main text, proofs of the five propositions, and figures).2. Experiments/ — 17 fully scripted experiment pipelines that build EduGrade-Bench v1 (41,800 records), run the seeded simulation calibrated to the real run's aggregates (what-if ranking, QWK 0.881 vs. 0.741 for multi-agent debate), and regenerate every table and figure in the paper.3. RealStudents/ — 5 scripts for the exploratory check on 120 real student responses (ASAP-SAS), plus the RealRun/ records of the real-call validation on the production stack, blind-audited by a different model family (quotation traceability 0.93, 97.5% blind-audit agreement, kappa = 0.95).4. README and REPRODUCIBILITY — step-by-step environment setup, execution order, expected outputs, and the mapping from each script to the paper's tables and figures.5. SHA256SUMS.txt — checksums of every file for integrity verification. Comments in all scripts are in English. Chinese strings that remain inside the code (answer templates, subject-name matchers, gateway concurrency messages, method labels) are functional payloads required to reproduce the published results byte-for-byte and are documented in REPRODUCIBILITY. This work was supported in part by the Key Field Research Project of Guangdong Universities under Grant 2025ZDZX1012.

提供机构:
Zenodo
创建时间:
2026-11-01
二维码
社区交流群
二维码
科研交流群
商业服务