henrichyusef/chemdfm-odor-lora

ChemDFM odor predictor (LoRA)

QLoRA adapter for OpenDFM/ChemDFM-v1.5-8B that predicts the odor descriptors of a molecule from its SMILES. Output groups: Primary, Sub.

Usage

The SMILES must be RDKit-canonical, and the prompt must match the training template exactly.

import torch
from rdkit import Chem
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True,
                         bnb_4bit_compute_dtype=torch.bfloat16)
base = AutoModelForCausalLM.from_pretrained("OpenDFM/ChemDFM-v1.5-8B", quantization_config=bnb, device_map="auto")
model = PeftModel.from_pretrained(base, "henrichyusef/chemdfm-odor-lora").eval()
tok = AutoTokenizer.from_pretrained("henrichyusef/chemdfm-odor-lora")

smiles = Chem.MolToSmiles(Chem.MolFromSmiles("CC(=O)OCC(C)C"))
prompt = "[Round 0]\nHuman: List the primary odors and sub-odors of the molecule CC(=O)OCC(C)C. Answer only with the odor labels.\nAssistant:".replace("CC(=O)OCC(C)C", smiles)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=200, do_sample=False)
print(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

odor_pipeline/vocab.json lists the label vocabulary (and merged aliases) used for parsing and scoring.