from typing import List, Dict, Optional from transformers import pipeline LABEL_POS = "POSITIVE" LABEL_NEG = "NEGATIVE" LABEL_NEU = "NEUTRAL" def load_pipeline(model_name: str = "distilbert-base-uncased-finetuned-sst-2-english"): """ Loads a local sentiment-analysis pipeline. NOTE: On first run, Transformers may download model weights to your machine. """ return pipeline("sentiment-analysis", model=model_name) def soften_to_neutral(pos_score: float, neg_score: float, margin: float) -> Optional[str]: """ If the top class wins by less than `margin`, return NEUTRAL. margin in [0, 0.5]. Larger => more neutral outputs. """ gap = abs(pos_score - neg_score) return LABEL_NEU if gap < margin else None def postprocess(raw_item: Dict, margin: float) -> Dict: label = raw_item["label"].upper() score = float(raw_item["score"]) pos = score if label == LABEL_POS else 1.0 - score # two-class trick neg = 1.0 - pos neutral = soften_to_neutral(pos, neg, margin) if neutral: return {"label": LABEL_NEU, "score": float(1.0 - abs(pos - neg))} # Return winning class + score if pos >= neg: return {"label": LABEL_POS, "score": pos} else: return {"label": LABEL_NEG, "score": neg} def analyze_batch(texts: List[str], pipe=None, neutral_margin: float = 0.15) -> List[Dict]: """ Runs local inference on a list of strings and applies neutral margin postprocessing. """ if pipe is None: pipe = load_pipeline() raw = pipe(texts, truncation=True) # HF pipeline returns single dict for single input; normalize to list if isinstance(raw, dict): raw = [raw] return [postprocess(item, neutral_margin) for item in raw]