Comparing Architecture Classification

Clean export package for the architecture benchmark experiment.

Contents:

• benchmark_main.py: runs the three-architecture benchmark
• create_eval_slice.py: builds balanced evaluation slices
• benchmark_config.json: benchmark settings
• Eval_Slices/: balanced 100-row evaluation slice used in the final fair run
• Results/BENCH_20260324_132323/: final fair run outputs
• SUMMARY.md: concise findings

Architectures benchmarked:

1. openai_only
2. ensemble_judge
3. chain_review

Final headline result:

• ensemble_judge performed best on the balanced 100-row evaluation slice.

Dataset Source

Original dataset: Mental Health Text Classification Dataset (Kaggle)

The full raw dataset is not included in this repository. This repository includes only:

• benchmark code
• a balanced 100-row evaluation slice derived from the dataset
• final benchmark outputs