silversuncn/eval-budget-robustness

Reproduction bundle for evaluation-budget robustness in low-resource text classification # Evaluation-Budget Robustness in Low-Resource Text Classification > **Evaluation-Budget Robustness in Low-Resource Text Classification** > Yaowen Sun ## Overview This repository contains a sanitized reproduction bundle for a low-resource text-classification study on how evaluation sample size affects empirical method ranking and false winner selection. Repository URL: . The bundle includes the formal result matrix, analysis script, cached summary files, paper figures, lightweight tests, and dependency notes needed to reproduce the reported decision-risk summaries. The formal matrix contains 6480 raw rows and the analysis pipeline produces 360 aggregated rows. The experiment covers five datasets, three probe methods, four training budgets, three training seeds, five numeric evaluation budgets, and deterministic full-anchor evidence. Primary paired inference is performed over decision cells keyed by dataset, training budget, training seed, and evaluation seed; each dataset/train-budget stratum has 21 primary paired decision cells. ## Repository Structure ```text . ├── README.md # Repository guide and reproduction notes ├── CITATION.cff # Minimal citation metadata ├── LICENSE # License text ├── requirements.txt # Minimal dependencies for analysis and tests ├── data/ │ ├── README.md # Data-file notes │ ├── formal_matrix_raw_results_20260726.csv │ │ # 6480-row formal result matrix │ ├── results_aggregated.csv # 360-row aggregated analysis output │ ├── phase4_analysis_20260726.json # Machine-readable summary and checks │ └── class_count_validation_20260727.json │ # Class-count validation summary ├── figures/ │ ├── false_selection_by_budget.png # False-selection figure │ ├── false_selection_by_budget.pdf # PDF version of false-selection figure │ ├── rank_stability_by_ …