Kirosama/medical-guardrail-mmbert-V4

Medical Guardrail (mmBERT) V4

This is an optimized, multi-label sequence classification model designed for medical routing and safety guardrails. It is built on top of jhu-clsp/mmBERT-base and fine-tuned for high-speed, enterprise-grade inference.

Architectures Included

This repository contains the model in three different formats to support various deployment environments:

• —Standard PyTorch: Available in the root directory.
• —ONNX (FP32): Available in the onnx/fp32 directory.
• —ONNX (INT8 Quantized): Available in the onnx/int8 directory (Recommended for CPU production).

Classes & Priority Order

The model outputs probabilities for 6 distinct classes. In production, we route based on the following strict priority order to ensure safety:

1. 1.Injection (Class 5) - Highest Priority (Instantly block prompt injection/jailbreaks)
2. 2.Emergency (Class 0) - (Route to human/emergency protocols)
3. 3.Malice_Abuse (Class 4) - (Block abusive language)
4. 4.Medical_Illogical (Class 2) - (Flag logically flawed medical queries)
5. 5.Medical_Logical (Class 1) - (Safe medical queries)
6. 6.General_Chat (Class 3) - Lowest Priority (Standard non-medical conversation)

Note on Outputs: Because this is a Multi-Label classification model, the model outputs raw logits. You must apply a Sigmoid activation function (e.g., torch.sigmoid() or scipy.special.expit()) to the outputs to get the final confidence probabilities. Do not use Softmax.

Usage

1. Standard PyTorch

Because the PyTorch files are in the root directory, you can load them using standard transformers.

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

repo_id = "Kirosama/medical-guardrail-mmbert-V4"
# fix_mistral_regex=True silences the known upstream regex warning
tokenizer = AutoTokenizer.from_pretrained(repo_id, fix_mistral_regex=True)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)

inputs = tokenizer("I have severe chest pain.", return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
    # Apply Sigmoid to get probabilities
    probabilities = torch.sigmoid(logits)[0].tolist()

2. ONNX (FP32)

To load the ONNX graph, switch from transformers to optimum and specify the subfolder.

from optimum.onnxruntime import ORTModelForSequenceClassification
from transformers import AutoTokenizer

repo_id = "Kirosama/medical-guardrail-mmbert-V4"
tokenizer = AutoTokenizer.from_pretrained(repo_id, fix_mistral_regex=True)

ort_fp32_model = ORTModelForSequenceClassification.from_pretrained(
    repo_id,
    subfolder="onnx/fp32"
)

3. ONNX (INT8 Quantized) - Production

This is the ultra-fast production model. You must specify both the subfolder and the custom file_name because the quantizer appends _quantized to the file.

from optimum.onnxruntime import ORTModelForSequenceClassification
from transformers import AutoTokenizer

repo_id = "Kirosama/medical-guardrail-mmbert-V4"
tokenizer = AutoTokenizer.from_pretrained(repo_id, fix_mistral_regex=True)

ort_int8_model = ORTModelForSequenceClassification.from_pretrained(
    repo_id,
    subfolder="onnx/int8",
    file_name="model_quantized.onnx"
)