reproduce/README.md

Reproduction guide

This directory contains the necessary information and assets to reproduce the results obtained during this Heretic run.

Local code

This system installed Heretic from a local directory or wheel. Uncommitted or experimental code may have been executed.

Reproducibility cannot be guaranteed in this environment.

Models

• Base model: google/gemma-4-E4B-it (Commit: fee6332)

Datasets

• Good prompts: mlabonne/harmless_alpaca (Commit: 02c6a92)
• Bad prompts: mlabonne/harmful_behaviors (Commit: 01cead0)
• Good evaluation prompts: mlabonne/harmless_alpaca (Commit: 02c6a92)
• Bad evaluation prompts: mlabonne/harmful_behaviors (Commit: 01cead0)

Selected trial

• Trial number: 88
• KL divergence: 0.007246
• Refusals: 42/100

System

• Python: 3.12.3 (CPython, GCC 13.3.0) [System]
• Operating system: Linux-6.8.0-111-generic-x86_64-with-glibc2.39 (x86_64)
• CPU: AMD EPYC 7713P 64-Core Processor

Accelerators

• CUDA: Detected 1 device(s) (31.37 GB total VRAM)
  • CUDA Version: 12.8
  • Driver Version: 580.159.03
• Devices:
  • CUDA 0: NVIDIA RTX PRO 4500 Blackwell (31.37 GB)

Environment

• Heretic: v1.3.0 (Origin: Local)
• PyTorch: 2.8.0+cu128
• Other dependencies: See requirements.txt.

Contents of this directory

• requirements.txt: The exact versions of all Python packages.
• config.toml: The exact configuration used, including the RNG seed.
• google--gemma-4-E4B-it.jsonl: The Optuna study journal containing the history of all trials.
• SHA256SUMS: Cryptographic hashes for all weight files.
• reproduce.json: A machine-readable file containing all reproducibility information.

How to reproduce

1. Ensure your system matches the specifications in the System section above. Exact reproducibility is only guaranteed if all aspects of your system are identical to the one the model was originally generated on.
2. Install the exact version of Heretic indicated in the Environment section above, from its original source.
3. Install the packages listed in requirements.txt: pip install -r requirements.txt
4. Install the correct version of PyTorch: pip install torch==2.8.0+cu128 --index-url https://download.pytorch.org/whl/cu128
5. Place the provided config.toml in your working directory.
6. Run Heretic without any additional arguments: heretic
7. Wait for the run to finish, then select trial 88 and export the model.
8. Verify that the weight files have been exactly reproduced by comparing their SHA-256 hashes against those in SHA256SUMS: sha256sum -c SHA256SUMS (or look at the hashes online if you uploaded to Hugging Face)

To use the included Optuna study journal google--gemma-4-E4B-it.jsonl, place it in the checkpoints directory (usually checkpoints/) before running Heretic.

This allows you to export other models from the Pareto front, or to run additional trials without having to re-run the stored trials.