| """Check Toxicity-Lens against Perspective scores you already logged, before you switch. |
| |
| Input: a CSV or JSONL file with a `text` column and one column per Perspective attribute you logged, named TOXICITY, INSULT, ... |
| (or perspective_TOXICITY, ...). For each attribute it prints rank agreement and, at your threshold, how many of Perspective's |
| flags this model keeps, how many extra texts it flags, and the threshold that flags the same share of texts as Perspective did. |
| Run: python compare_with_perspective.py logged.csv [--model abacusai/Toxicity-Lens] [--backend onnx] [--threshold 0.7] |
| """ |
| import argparse |
| import csv |
| import json |
|
|
| import numpy as np |
|
|
| from server import ATTRS, Scorer |
|
|
|
|
| def rank(x): |
| r = np.empty(len(x)) |
| r[np.argsort(x, kind="stable")] = np.arange(len(x)) |
| return r |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("file") |
| ap.add_argument("--model", default="abacusai/Toxicity-Lens") |
| ap.add_argument("--backend", default="onnx", choices=["onnx", "torch"]) |
| ap.add_argument("--threshold", type=float, default=0.5, help="the threshold you use with Perspective") |
| a = ap.parse_args() |
| with open(a.file, encoding="utf-8", newline="") as f: |
| rows = [json.loads(line) for line in f if line.strip()] if a.file.endswith(".jsonl") else list(csv.DictReader(f)) |
| scorer = Scorer(a.model, a.backend) |
| ours = [s for i in range(0, len(rows), 256) for s, _ in scorer([str(r["text"]) for r in rows[i:i + 256]])] |
| t = a.threshold |
| print(f"{len(rows)} texts, threshold {t}\n{'attribute':<18}{'n':>7}{'spearman':>10}{'persp flags':>13}{'our flags':>11}" |
| f"{'kept':>8}{'extra':>8}{'same-rate threshold':>21}") |
| for attr in ATTRS: |
| col = next((c for c in (attr, f"perspective_{attr}", attr.lower(), f"perspective_{attr.lower()}") if c in rows[0]), None) |
| keep = [i for i, r in enumerate(rows) if col and r.get(col) not in (None, "")] |
| if len(keep) < 2: |
| continue |
| p = np.array([float(rows[i][col]) for i in keep]) |
| o = np.array([ours[i][attr] for i in keep]) |
| flag = p >= t |
| rho = np.corrcoef(rank(p), rank(o))[0, 1] |
| kept = f"{(o[flag] >= t).mean():.1%}" if flag.any() else "-" |
| extra = f"{(o[~flag] >= t).mean():.1%}" if (~flag).any() else "-" |
| same = f"{np.quantile(o, 1 - flag.mean()):.3f}" if 0 < flag.mean() < 1 else "-" |
| print(f"{attr:<18}{len(keep):>7}{rho:>10.3f}{flag.mean():>13.1%}{(o >= t).mean():>11.1%}{kept:>8}{extra:>8}{same:>21}") |
| print("\nkept = share of Perspective's flags this model also flags; extra = share of texts Perspective left below the threshold " |
| "that this model flags; same-rate threshold = the threshold at which this model flags as many texts as Perspective did.") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
