Dates Generator -- Research-Backed Submission

Conditional date generation given (DOW, MON, LEAP, DEC) conditions, with four research-backed generative models trained from scratch on a Google Colab GPU runtime.

Run on Google Colab (training, evaluation, full walkthrough)

1. Click the Open In Colab badge above (or open notebook.ipynb on Colab manually via File -> Open notebook -> GitHub tab).
2. Runtime -> Change runtime type -> GPU (T4).
3. Run all cells. The first cell clones this public repo into /content/dates-generator, !pip install -r requirements.txt, then the notebook trains all four models (~60-90 min total on T4), evaluates, runs the CFG ablation, and writes a predictions.txt from the example input.

Run inference locally (matches assignment CLI)

conda env create -f environment.yml
conda activate dates-generator
cd model
python predict.py -i ../data/example_input.txt -o ../predictions.txt

Pick a different model with --model {cvae,acgan,maskgit,mdlm}. Default: cvae.

Retrain from scratch

python model/train.py --data data/data.txt --out model/weights \
    --models cvae acgan maskgit mdlm

Run tests

pytest

Layout

• model/predict.py -- CLI entry point
• model/train.py -- training orchestrator
• model/training.py -- per-model trainers
• model/tokenizer.py, model/data.py, model/metrics.py, model/evaluation.py
• model/models/{cvae,acgan,maskgit,mdlm}.py -- the four model implementations
• model/weights/{cvae,acgan,maskgit,mdlm}.pt -- trained weights
• model/weights_legacy/ -- previous submission's weights (kept for comparison)
• tests/ -- pytest tests
• notebook.ipynb -- Colab walkthrough
• report/ -- assignment write-up
• scripts/run_final_eval.py, scripts/run_cfg_ablation.py
• results/ -- CSR tables, diversity, figures

Models

#	Model	Class	Citation
1	CVAE + CFG dropout	In-course	Sohn 2015; Ho & Salimans 2022
2	AC-GAN hybrid MLE + WGAN-GP + projection-D	In-course (required GAN)	Odena 2017; Che 2017; Gulrajani 2017; Miyato/Koyama 2018
3	MaskGIT	Out-of-course	Chang et al. CVPR 2022
4	MDLM + CFG + joint head	Out-of-course	Sahoo et al. NeurIPS 2024

See docs/superpowers/specs/2026-05-17-dates-generator-research-backed-redesign-design.md for the design rationale.

Reproducibility

Seed 0 everywhere (torch, numpy, random). Dates range [1-1-1800, 31-12-2200] per the assignment constraint.