knowledgator/opir-multitask-large-v1.0 · Hugging Face

What it is

knowledgator/opir-multitask-large-v1.0 — the English, highest-accuracy multi-task checkpoint in Knowledgator's Opir GLiClass guardrail family for real-time LLM safety filtering. The page is the model card: it gives the actual integration path, the pipeline modes, and the threshold guidance the summary omits. The Opir paper reportedly ships three companion checkpoints for other deployment profiles (not named on this page).

How it works

It's an encoder model that jointly encodes input text and candidate labels via GLiClass, rather than generating a verdict token by token. You supply candidate labels at inference time, so policy changes need no retraining. Two modes matter:

• classification_type="single-label" → binary safe/unsafe.
• classification_type="multi-label" → taxonomy, toxicity, jailbreak, or custom policy labels, scored independently against a threshold.

Labels can be plain strings, dot-notation, hierarchical dicts, or carry a prompt= task instruction. For prompt-response evaluation, the page's approach is to serialize the turn into one text field (f"Prompt: {prompt}\nResponse: {response}") and classify labels like ["safe response", "unsafe response", "refusal", "compliance"] — GLiClass has no native pair input.

What to do

pip install gliclass transformers

Binary decision:

from gliclass import GLiClassModel, ZeroShotClassificationPipeline
from transformers import AutoTokenizer
MODEL_ID = "knowledgator/opir-multitask-large-v1.0"
model = GLiClassModel.from_pretrained(MODEL_ID)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
binary_classifier = ZeroShotClassificationPipeline(
    model=model, tokenizer=tokenizer,
    classification_type="single-label", device="cuda:0")
result = binary_classifier(text, ["safe", "unsafe"])[0]
print(max(result, key=lambda x: x["score"]))

Multi-label runs the same pipeline with classification_type="multi-label" and a threshold= arg. The page lists concrete label sets for the 16 top-level categories, toxicity (6 labels), and jailbreak/prompt-injection (10 labels including instruction hierarchy attack, indirect prompt injection, secret or context exfiltration, tool and connector abuse).

Caveats

• The paper's default threshold=0.5 is explicitly a starting point — the card says calibrate thresholds on representative traffic before production.
• Label descriptions are conditional: "include descriptions if your GLiClass version/configuration supports them" — feature availability depends on your installed version.
• device="cuda:0" in examples; the card notes CPU is possible but this is the large/highest-accuracy checkpoint, so weigh latency for real-time use.
• Benign-sensitive contrast categories (defensive cybersecurity, counterspeech, harm prevention, appropriate refusal) exist to cut over-refusal, but you must actually include those labels in your schema to benefit.