Spaces:
Running on Zero
Running on Zero
Download ml_classifier.py from VanillaVector/Hybrid-Comment-Moderation: direct link, hf CLI and curl.
- Browser
- Download file 1.94 kB
-
https://huggingface.co/spaces/VanillaVector/Hybrid-Comment-Moderation/resolve/main/ml_classifier.py
- Command line
-
hf download hf://spaces/VanillaVector/Hybrid-Comment-Moderation/ml_classifier.py
-
curl -L -o ml_classifier.py https://huggingface.co/spaces/VanillaVector/Hybrid-Comment-Moderation/resolve/main/ml_classifier.py
1.94 kB
| """ | |
| ml_classifier.py — DistilBERT toxic-comment classifier (Stage 1). | |
| Runs on CPU only — no GPU quota consumed. Inference takes ~7 ms | |
| per comment, so it is available to every visitor (including guests). | |
| """ | |
| from typing import Tuple | |
| from transformers import pipeline as hf_pipeline | |
| from config import ( | |
| CLASSIFIER_MODEL_NAME, | |
| CONFIDENCE_HIGH, | |
| CONFIDENCE_LOW, | |
| DECISION_BLOCK, | |
| DECISION_ALLOW, | |
| ) | |
| class MLClassifier: | |
| """Wrapper around a HuggingFace text-classification pipeline.""" | |
| def __init__(self, model_name: str = CLASSIFIER_MODEL_NAME): | |
| self._pipe = hf_pipeline( | |
| task="text-classification", | |
| model=model_name, | |
| tokenizer=model_name, | |
| framework="pt", | |
| device=-1, # ← force CPU (no GPU quota) | |
| truncation=True, | |
| max_length=512, | |
| ) | |
| self._toxic_labels = {"toxic"} | |
| print(f"✅ ML Classifier loaded on CPU: {model_name}") | |
| def classify(self, text: str) -> Tuple[str, float, bool]: | |
| """ | |
| Returns (decision, confidence, is_uncertain). | |
| decision : DECISION_BLOCK or DECISION_ALLOW | |
| confidence : float in [0, 1] | |
| is_uncertain : True when confidence is in the grey zone | |
| """ | |
| result = self._pipe(text)[0] | |
| return self._interpret(result) | |
| def _interpret(self, result: dict) -> Tuple[str, float, bool]: | |
| label = result["label"].lower() | |
| score = result["score"] | |
| toxic_prob = score if label in self._toxic_labels else (1.0 - score) | |
| if toxic_prob >= CONFIDENCE_HIGH: | |
| return (DECISION_BLOCK, toxic_prob, False) | |
| elif toxic_prob <= CONFIDENCE_LOW: | |
| return (DECISION_ALLOW, 1.0 - toxic_prob, False) | |
| else: | |
| tentative = DECISION_BLOCK if toxic_prob >= 0.5 else DECISION_ALLOW | |
| return (tentative, max(toxic_prob, 1.0 - toxic_prob), True) |