Code and results archive for "Enhancing Routing Fairness in Hybrid Scoring With Group-Conditional Coverage"

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

Code, engine score probabilities, and result tables reproducing every number, table, and figure in the manuscript "Enhancing Routing Fairness in Hybrid Scoring With Group-Conditional Coverage". The study audits whether hybrid scoring systems, which route responses an automated engine cannot score confidently to human raters, protect student groups equally. Routing is expressed as a split conformal prediction guarantee, and one pooled calibration threshold is compared against one threshold per group across three scoring engines (a fine-tuned RoBERTa and two zero-shot large language models), two corpora with student demographics (PERSUADE 2.0 and ELLIPSE), six groupings, two score functions, six risk levels, and 20 calibration splits. Contents: the five engine scripts at the archive root; analysis/ with the calibration, robustness, and figure scripts, the result tables they write, and the figures; persuade_data/ and ellipse_data/ with the engine score probabilities, the frozen calibration splits, and provenance notes for each corpus; NUMBERS.md mapping every number in the paper to the command that produced it; MANIFEST.txt with a SHA-256 for every file. Neither corpus is redistributed here. PERSUADE 2.0 and ELLIPSE are obtained from their official repositories, which the README links; ellipse_prep.py rebuilds the ELLIPSE analysis file, which is excluded because it carries essay text. The included probability files contain response identifiers and model score probabilities only. The directory layout mirrors the working layout because each script resolves its inputs relative to its own location; run the scripts where they sit. Licensing: the code is the authors' own work and is released under the MIT License (LICENSE_CODE.txt). The score probability files are derivatives of two corpora released under CC BY-NC-SA 4.0 and inherit that license (LICENSE_DATA.txt), which is why the record as a whole carries the more restrictive of the two.

Links and sources

Comments