Repository navigation
Selected adapters from the completed molecular-origami post-training experiment.
Included are the supervised LoRA checkpoint, RL seed 17, RL seed 29, populated model cards, tokenizer files, upstream notices, and SHA-256 checksums. Base Qwen3-8B weights are not included. The repository's adapters/manifest.json records archive and member digests; results/TRAINING_FROZEN.json binds the selected tensors to the original run.
Install the archives with scripts/install_adapters.py and use the custom constrained decoder described in docs/REPRODUCIBILITY.md. Ordinary chat generation does not reproduce this policy.
All final model evaluations used FP32, but the original RL updates used BF16 base computations with a documented probability-consistency limitation. The proposed full-precision RL follow-up has not been run. This is a computational secondary-structure study with no laboratory validation; RNAinverse remains the stronger practical baseline on the reported exact-fold endpoint.
The source, data, baselines, candidate ledgers, numerical amendments, and illustrated essay are in this repository. The essay is published at https://aryasomu.com/writing/molecular-origami.html.