C-MORAL-Mistral-GRPO

C-MORAL Mistral GRPO Adapters

This repository contains LoRA adapters released for C-MORAL:

C-MORAL: Controllable Multi-Objective Molecular Optimization with Reinforcement Alignment for LLMs

These adapters are trained on top of:

• mistralai/Mistral-7B-Instruct-v0.3

using:

• GRPO

for controllable multi-objective molecular optimization.

Available Task Subfolders

Each task is stored as a separate subfolder in this Hugging Face repository.

• abmp: amp+bbbp+mutag+plogp
• acep: amp+carc+herg+plogp
• bcmq: bbbp+carc+mutag+qed
• bdeq: bbbp+drd2+herg+qed
• bdpq: bbbp+drd2+qed+plogp
• bpq: bbbp+plogp+qed
• cde: carc+drd2+herg
• dhmq: drd2+hia+mutag+qed
• elq: herg+liv+qed
• hlmpq: hia+liv+mutag+plogp+qed

Usage

Load a task-specific adapter with PEFT by setting subfolder to the desired task name.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "mistralai/Mistral-7B-Instruct-v0.3"
adapter_repo = "Rwigle/C-MORAL-Mistral-GRPO"
task_subfolder = "bpq"  # change to abmp / elq / hlmpq / ...

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(base_model_id)
model = PeftModel.from_pretrained(model, adapter_repo, subfolder=task_subfolder)

Method

• Base model: mistralai/Mistral-7B-Instruct-v0.3
• Adapter type: LoRA
• Training algorithm: GRPO
• Domain: multi-objective molecular optimization

Project

• GitHub: https://github.com/Rwigie/C-MORAL

Citation

If you use these adapters, please cite:

C-MORAL: Controllable Multi-Objective Molecular Optimization with
Reinforcement Alignment for LLMs