Public reproducibility repository for PSCFM algorithms and experiments. The repository contains analytical validation, image-benchmark workflows, parameter-matched real-valued controls, and CH-SIMS v2.0 multimodal-conflict experiments.
- GitHub repository: https://github.com/xwxkj/PSCFM
- Software archive (Zenodo v1.2.2): https://doi.org/10.5281/zenodo.21941667
- Processed-results archive (Zenodo v1.2.2): https://doi.org/10.5281/zenodo.21941698
The GitHub repository is the browsable source tree. The Zenodo software DOI identifies the immutable version used for the study, and the separate results DOI identifies the corresponding processed experimental outputs.
src/,run_all.py: analytical checks and CIFAR-C, PACS, and Office-Home workflows.matched_baselines/: parameter-matched Real Context Gate and higher-capacity Real Context MLP comparisons.chsims_v2/: CH-SIMS v2.0 feature validation, multimodal-conflict experiments, selective prediction, and aggregation.configs/: compact and full example configurations.scripts/: dataset verification, multi-seed execution, aggregation, and environment capture.
conda env create -f environment.yml
conda activate pscfm_corepython run_all.py --quick_test --output_dir outputs/quick
python matched_baselines/scripts/quick_test_matched_baselines.py
bash chsims_v2/scripts/quick_test.shOriginal third-party datasets are not distributed. Obtain CIFAR-10-C, CIFAR-100-C, PACS, Office-Home, and CH-SIMS v2.0 from their official providers. Dataset references and expected directory layouts are documented in the workflow directories.
Formal workflows use seeds 2024–2028. Configuration files, dependency specifications, dataset checks, and aggregation utilities are included. For compatibility with archived outputs, some command-line identifiers and CSV column names retain the historical token pscm; these identifiers refer to PSCFM and do not alter the implemented method.
MIT License. Third-party datasets and external baseline implementations remain subject to their own terms.