Ministral-DrugDetector-14B-All9

A LoRA fine-tune of Ministral-3-14B-Instruct-2512 for multi-class substance use detection in clinical notes. Given a medical note, the model predicts illicit or harmful use across 9 substance classes and classifies each as current, historical, or unknown — all in a single forward pass.

Substances Detected

Class	Label
Methamphetamine	Illicit use
Fentanyl	Illicit use
Injection Drug Use (IDU)	Any route
Heroin	Illicit use
Prescription Opioid Misuse	Non-prescribed / abusive context
Cocaine	Illicit use
Alcohol	Harmful use (AUD, intoxication, withdrawal)
Cannabis	Illicit / recreational use
Benzodiazepines	Misuse / non-prescribed

Output Format

Methamphetamine Illicit Use: True/False/Unknown
Fentanyl Illicit Use: True/False/Unknown
Injection Drug Use: True/False/Unknown
Heroin Illicit Use: True/False/Unknown
Prescription Opioid Misuse: True/False/Unknown
Cocaine Illicit Use: True/False/Unknown
Alcohol Harmful Use: True/False/Unknown
Cannabis Illicit Use: True/False/Unknown
Benzodiazepine Misuse: True/False/Unknown
Methamphetamine Temporal Status: Current/Historical/Unknown/N/A
Fentanyl Temporal Status: Current/Historical/Unknown/N/A
Injection Drug Use Temporal Status: Current/Historical/Unknown/N/A
Heroin Temporal Status: Current/Historical/Unknown/N/A
Prescription Opioid Temporal Status: Current/Historical/Unknown/N/A
Cocaine Temporal Status: Current/Historical/Unknown/N/A
Alcohol Temporal Status: Current/Historical/Unknown/N/A
Cannabis Temporal Status: Current/Historical/Unknown/N/A
Benzodiazepine Temporal Status: Current/Historical/Unknown/N/A

Performance (held-out test set, n=187 clinical notes)

Illicit Use Detection (F1, positive class = True)

Substance	Precision	Recall	F1
Methamphetamine	1.000	0.963	0.981
Fentanyl	0.971	0.971	0.971
Rx Opioid Misuse	1.000	0.917	0.957
Cannabis	0.971	0.943	0.957
Benzodiazepine Misuse	0.967	0.935	0.951
Heroin	1.000	0.897	0.945
Injection Drug Use	1.000	0.886	0.939
Cocaine	1.000	0.844	0.915
Alcohol (harmful)	1.000	0.795	0.886

Temporal Classification (accuracy among True cases)

Substance	N	Accuracy
Benzodiazepine Misuse	24	0.833
Fentanyl	35	0.800
Methamphetamine	41	0.756
Heroin	22	0.773
Cannabis	25	0.760
Rx Opioid Misuse	31	0.742
Injection Drug Use	27	0.630
Cocaine	24	0.583
Alcohol (harmful)	27	0.556

Training

• Base model: mistralai/Ministral-3-14B-Instruct-2512
• Method: LoRA (r=16, alpha=32, dropout=0.05, targets: q/k/v/o_proj)
• Training data: 495 manually annotated clinical notes (UCLA Health, de-identified)
• Epochs: 3 | Batch size: 1 (grad accum 2) | LR: 2e-4
• Trainable parameters: 22.8M (0.16% of 14B)
• Hardware: 1× NVIDIA RTX A6000 (48GB)

Usage

from transformers import AutoTokenizer
from peft import PeftModel
import torch

try:
    from transformers import Mistral3ForConditionalGeneration
    model = Mistral3ForConditionalGeneration.from_pretrained(
        "mistralai/Ministral-3-14B-Instruct-2512",
        torch_dtype=torch.bfloat16, device_map="auto"
    )
except ImportError:
    from transformers import AutoModelForCausalLM
    model = AutoModelForCausalLM.from_pretrained(
        "mistralai/Ministral-3-14B-Instruct-2512",
        torch_dtype=torch.bfloat16, device_map="auto"
    )

model = PeftModel.from_pretrained(model, "fabriceyhc/Ministral-DrugDetector-14B-All9")
model.eval()
tokenizer = AutoTokenizer.from_pretrained("fabriceyhc/Ministral-DrugDetector-14B-All9")

note = "Patient presents with polysubstance use disorder. History of IV heroin use, now on MAT. Recent urine tox positive for methamphetamine."

prompt = f"""<s>[INST] ### Task Description:
Please carefully review the following medical note and identify illicit or harmful substance use.

### The medical note to evaluate:
{note}

### Answer:
 [/INST]"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=300, do_sample=False)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Intended Use

Designed for NLP annotation of de-identified clinical notes in research settings. Not intended for clinical decision-making.

Framework versions

• PEFT 0.18.1