ABI-CDMs

Amortized Bayesian inference for conflict diffusion models, accompanying the manuscript No Single Model Fits All: Conflict Decision-Making Models Vary More Across Studies Than Across Tasks.

The repository contains the release-ready pipeline for four primary models—DDM, DMC, SSP, and DSTP—plus the constrained and dRiftDM-aligned specifications reported in the supplementary material. Notebook demonstrations, ad hoc tests, and unrelated experimental models are excluded.

Repository layout

ABI-CDMs/
├── data/                       # Nine retained raw CSV files
├── checkpoints/                # Downloaded pretrained estimators
├── figures/
│   ├── main/                   # Canonical Figures 1–5 (PNG only)
│   └── supplement/             # Generated supplementary figures
├── nsbi_module/                # Reusable inference and analysis library
├── results/
│   ├── intermediate/           # Generated HDF5, pickle, CSV, and RDS files
│   └── tables/                 # Generated result tables
├── scripts/                    # Pipeline in execution order
├── docs/DATA_SOURCES.md        # Raw and analysis sample documentation
└── SOURCE_PROVENANCE.md        # Mapping to the development repository

All Python paths are derived from the repository root. Run the commands below from the repository root; changing into individual script directories is not required.

Installation

Python 3.10–3.12 and R 4.0 or newer are recommended. A GPU is optional for inference but strongly recommended for training. Inkscape is required only for the optional SVG-to-PDF export and serves as the final PNG-rendering fallback for Figure 4.

conda env create -f environment.yml
conda activate abi-cdms
pip install -e .

For the factor and reliability analyses, install the R packages used by the files in scripts/07_parameter_analysis/, including tidyverse, psych, GPArotation, brms, tidybayes, posterior, cmdstanr, and svglite.

Data scope

The release contains nine raw CSV files:

• eight studies used in the cross-sectional analysis;
• Raw-file participant counts differ from final analysis counts after task selection, session selection, and quality filtering. See DATA_SOURCES.md for the authoritative distinction.

Pretrained checkpoints

Download the archived checkpoints from Zenodo, then extract them into checkpoints/:

checkpoints/
├── DDM/
├── DMC/
├── SSP/
├── DSTP/
└── driftdm_dmc/

This archive covers the four primary models and the six-parameter dRiftDM-aligned DMC. The reduced DMC, SSP, and DSTP checkpoints used for Figure S7, and the seven-parameter variable-start DMC used by the current Figure S8 pipeline, are not in the current archive; use the training commands below to rebuild them.

Reproduce the analysis

1. Preprocess raw data

python scripts/01_preprocessing/prepare_datasets.py

Creates results/intermediate/datasets_cross_sectional.h5 and datasets_retest.h5.

2. Train estimators (optional)

Skip this step when using the pretrained checkpoints.

python scripts/02_training/train_ddm.py
python scripts/02_training/train_dmc.py
python scripts/02_training/train_ssp.py
python scripts/02_training/train_dstp.py

The supplementary constrained specifications have separate, semantically named entry points:

python scripts/02_training/train_dmc_fixed_shape.py
python scripts/02_training/train_ssp_fixed_ratio.py
python scripts/02_training/train_dstp_fixed_ratio.py
python scripts/02_training/train_driftdm_aligned_dmc.py
python scripts/02_training/train_driftdm_aligned_dmc_variable_start.py

The first three reduce weakly identifiable parameter combinations. The six-parameter dRiftDM-aligned model fixes the automatic-activation shape and centers the starting point; the seven-parameter version additionally estimates symmetric starting-point variability. They are intentionally kept as distinct model registrations and checkpoint directories.

3. Fit models and summarize predictions

python scripts/03_fitting/fit_core_models.py
python scripts/03_fitting/fit_extended_dmc.py
python scripts/03_fitting/summarize_core_fits.py
python scripts/03_fitting/summarize_extended_dmc.py

4. Run validation analyses

python scripts/04_validation/figure_s01_s03_parameter_recovery.py
python scripts/04_validation/figure_s01_s04_model_recovery.py
python scripts/04_validation/figure_s10_parameter_mapping.py
python scripts/04_validation/figure_s07_reduced_model_recovery.py

5. Generate model-comparison and PPC figures

python scripts/05_model_comparison/figure_02_model_comparison.py
python scripts/06_ppc/generate_ppc_data.py
python scripts/06_ppc/figure_03_posterior_predictive_checks.py
python scripts/06_ppc/figure_s05_caf.py

6. Estimate factors and generate reliability figures

Render scripts/07_parameter_analysis/estimate_factor_scores.Rmd, then run the reliability models before generating the final figures:

Rscript scripts/07_parameter_analysis/fit_reliability_models.R
python scripts/07_parameter_analysis/figure_04_latent_factors.py
python scripts/07_parameter_analysis/figure_05_factor_space.py
Rscript scripts/07_parameter_analysis/figure_s15_retest_icc.R
python scripts/07_parameter_analysis/figure_s16_representational_similarity.py

The Bayesian reliability step is computationally expensive and caches its models in results/intermediate/.

7. Generate robustness figures

python scripts/08_supplementary/figure_s17_rmse_scaling.py
python scripts/08_supplementary/figure_s18_ppc_component_metrics.py
python scripts/08_supplementary/figure_s19_model_metric_comparison.py
python scripts/08_supplementary/figure_s09_dstp_vs_dmc.py

Canonical manuscript figures

Figure	Generator	Published files
Figure 1	Design asset; no analysis generator	figures/main/figure_01_workflow.png
Figure 2	scripts/05_model_comparison/figure_02_model_comparison.py	figure_02_model_comparison.png
Figure 3	scripts/06_ppc/figure_03_posterior_predictive_checks.py	figure_03_posterior_predictive_checks.png
Figure 4	scripts/07_parameter_analysis/figure_04_latent_factors.py	figure_04_latent_factors.png
Figure 5	scripts/07_parameter_analysis/figure_05_factor_space.py	figure_05_factor_space.png

To export the five PNG masters as PDFs:

python scripts/09_export/export_main_figure_pdfs.py

License and citation

The code is licensed under AGPL-3.0; see LICENSE. If you use this release, cite the accompanying paper and the archived software record:

Pan, W., Wang, J., Oberauer, K., & Hu, C.-P. (2026). No Single Model Fits All: Conflict Decision-Making Models Vary More Across Studies Than Across Tasks.

Checkpoint archive: 10.5281/zenodo.21623907.