ACCEPT (0): one self-contained ratio, unit-rate, or percentage calculation.
REJECT (1): anything outside that scope, including prompt injection, wrappers, multiple questions, and requests for extra output.

Unlike a generative guard, this model exposes only classification logits. It is intended as a small, fail-closed gate in front of a downstream solver.

Try the live Hugging Face Space or see the source repository.

Evaluation

On the held-out test set (6,237 items), the training run reported:

Metric	Result
Accuracy	99.90%
Macro F1	0.9990
Held-out attack success rate	0.0%
False-reject rate	0.12% (6 items)

In a separate 144-item adversarial comparison, the adapter's overall attack-success rate was 3.5%, versus 63.2% for the prompt-only baseline used in that experiment. These results are specific to the repository's synthetic policy and evaluation setup; they are not a general safety guarantee.

Usage

python

import torch
from peft import PeftModel
from transformers import AutoModelForSequenceClassification, AutoTokenizer

adapter_id = "nishegde/ratio-filter-qwen3-1.7b-lora"
base_id = "Qwen/Qwen3-1.7B"
id2label = {0: "ACCEPT", 1: "REJECT"}
label2id = {label: idx for idx, label in id2label.items()}

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

base = AutoModelForSequenceClassification.from_pretrained(
    base_id,
    revision="70d244cc86ccca08cf5af4e1e306ecf908b1ad5e",
    num_labels=2,
    id2label=id2label,
    label2id=label2id,
    torch_dtype=torch.float32,
)
base.config.pad_token_id = tokenizer.pad_token_id
model = PeftModel.from_pretrained(base, adapter_id).eval()

text = "A car travels 300 miles on 10 gallons. What is its mileage?"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=1024)
with torch.inference_mode():
    probabilities = torch.softmax(model(**inputs).logits[0].float(), dim=-1)

print(id2label[int(probabilities.argmax())], probabilities.tolist())

Limitations

This is a narrow English-language policy classifier, not a general moderation or safety model.
Inputs outside its training distribution, other languages, heavy obfuscation, or very long prompts may be misclassified.
The base model must be downloaded separately; this repository contains the LoRA adapter, tokenizer, and trained classification head.
Treat the score threshold as an application decision and test on your own distribution before production use.
Lowering the rejection threshold makes the gate more fail-closed; raising it permits more inputs.

License

The base model declares Apache-2.0. No separate license has been specified for this adapter; users are responsible for reviewing the base model terms and obtaining any permissions needed for their use.