PlasmidGPT-GRPO

PlasmidGPT-GRPO

GRPO reinforcement-learning fine-tune of PlasmidGPT, trained against a multi-component biological reward (functional annotations, length prior, repeat penalty, cassette ordering). Camera-ready model for the ICML 2026 paper Effects of Structural Reward Shaping on Biophysical Properties in RL-Trained Plasmid Generators.

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("UCL-CSSB/PlasmidGPT-GRPO")
tokenizer = AutoTokenizer.from_pretrained("UCL-CSSB/PlasmidGPT-GRPO")

input_ids = tokenizer("ATG", return_tensors="pt").input_ids
outputs = model.generate(input_ids, max_new_tokens=512, do_sample=True, temperature=1.0)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Recommended sampling: T=1.0 for direct generation, T=1.15 for rejection sampling (per the paper).

Citation

@inproceedings{thiel2026plasmidrl,
  title     = {Effects of Structural Reward Shaping on Biophysical Properties in {RL}-Trained Plasmid Generators},
  author    = {Thiel, McClain and Cunningham, Angus G. and Barnes, Chris P.},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026}
}