| """DignityCheck eval — solo clasificadores (rápido). Tasa de falsos positivos sobre frases no-odio.""" |
| import json |
| import pandas as pd, torch |
| from transformers import AutoTokenizer, AutoModelForSequenceClassification |
| DEV = "mps" if torch.backends.mps.is_available() else "cpu" |
| df = pd.read_csv("ic/public/dignityai/dignitycheck/dignitycheck_v0.1.csv"); texts = df.texto.tolist() |
| @torch.no_grad() |
| def lg(tok, mod, bs=32): |
| o = [] |
| for i in range(0, len(texts), bs): |
| enc = tok(texts[i:i+bs], truncation=True, max_length=96, padding=True, return_tensors="pt").to(DEV) |
| o.append(mod(**enc).logits.float().cpu()) |
| return torch.cat(o) |
| def load(r): return AutoTokenizer.from_pretrained(r), AutoModelForSequenceClassification.from_pretrained(r).to(DEV).eval() |
| tk, md = load("piuba-bigdata/beto-contextualized-hate-speech"); S = torch.sigmoid(lg(tk, md)).numpy(); df["uba"] = S[:, [1,2,3,4,6,7]].max(1) |
| tk, md = load("pysentimiento/robertuito-hate-speech"); df["robertuito"] = torch.sigmoid(lg(tk, md)[:, 0]).numpy() |
| tk, md = load("antonn-dromundo/SinOdio-BETO-HateSpeech-Detector-v3"); df["sinodio"] = torch.softmax(lg(tk, md), -1)[:, 1].numpy() |
| MOD = ["uba", "robertuito", "sinodio"] |
| res = {"n": len(df), "umbral": 0.5, |
| "falsos_positivos_pct": {m: round(float((df[m] > 0.5).mean()) * 100, 1) for m in MOD}, |
| "por_grupo": {m: {g: round(float((s[m] > 0.5).mean()) * 100, 1) for g, s in df.groupby("grupo")} for m in MOD}, |
| "por_tipo": {m: {t: round(float((s[m] > 0.5).mean()) * 100, 1) for t, s in df.groupby("tipo")} for m in MOD}} |
| df.to_csv("dignitycheck_scores.csv", index=False) |
| json.dump(res, open("dignitycheck_eval.json", "w"), ensure_ascii=False, indent=2) |
| print(json.dumps(res, ensure_ascii=False, indent=1)) |
|
|
