kaiwei123/modernbert_par_2_jaur_1

MANTIS ModernBERT Command Router

This model classifies terminal commands for the MANTIS terminal simulator. It assigns each command to one of three routing categories:

Model output	MANTIS label	Intended meaning
LABEL_0	read	Primarily observes or retrieves system state
LABEL_1	write	May modify system state
LABEL_2	rejection	Should not be handled as an accepted command

The semantic label mapping is implemented by MANTIS because the model configuration retains the generic LABEL_0–LABEL_2 names.

Model Details

• Architecture: ModernBertForSequenceClassification
• Task: Single-label, three-class sequence classification
• Hidden size: 768
• Transformer layers: 22
• Attention heads: 12
• Maximum architectural context: 8,192 tokens
• MANTIS inference length: 256 tokens
• Tensor type: Float32
• Library: Hugging Face Transformers

Intended Use

The model is intended to route terminal commands inside the MANTIS simulated Ubuntu environment. MANTIS uses the prediction to choose between read-only handling, state-changing handling, and command rejection.

It is not intended to:

• authorize commands on a real operating system;
• serve as an independent security control;
• detect all malicious shell commands;
• replace sandboxing, access controls, or human review.

Usage

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "kaiwei123/modernbert_par_2_jaur_1"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

labels = {
    0: "read",
    1: "write",
    2: "rejection",
}

command = "find /tmp -type f -delete"

inputs = tokenizer(
    command,
    truncation=True,
    max_length=256,
    padding="max_length",
    return_tensors="pt",
)

with torch.no_grad():
    logits = model(**inputs).logits

prediction = int(logits.argmax(dim=-1).item())
print(labels[prediction])

To download the model locally:

hf download kaiwei123/modernbert_par_2_jaur_1 \
  --local-dir model/modernbert_par_2_jaur_1

Training Data

The associated routing dataset is available at:

• kaiwei123/routing-dataset

Limitations

• Commands can combine read and write operations, making a single label ambiguous.
• Shell aliases, scripts, quoting, pipes, substitutions, and obfuscation may alter command behavior.
• Behavior can depend on the operating system, installed programs, permissions, and current state.
• Inputs longer than 256 tokens are truncated by the MANTIS integration.
• Commands unlike the training data may be misclassified.
• A read prediction must not be treated as proof that execution is safe.

Security Considerations

This classifier is a routing component, not a security boundary. Applications should validate commands independently and execute untrusted input only inside an appropriately isolated environment.

Particular attention should be given to false read predictions because apparently read-only commands can invoke subprocesses or cause side effects through options, shell expansion, redirection, or command substitution.