it-no-bio-20251014-t21

Slur reclamation binary classifier
Task: LGBTQ+ reclamation vs non-reclamation use of harmful words on social media text.

Trial timestamp (UTC): 2025-10-14 11:10:28

Data case: it

Configuration (trial hyperparameters)

Model: Alibaba-NLP/gte-multilingual-base

Hyperparameter	Value
LANGUAGES	it
LR	2e-05
EPOCHS	3
MAX_LENGTH	256
USE_BIO	False
USE_LANG_TOKEN	False
GATED_BIO	False
FOCAL_LOSS	True
FOCAL_GAMMA	2.5
USE_SAMPLER	True
R_DROP	True
R_KL_ALPHA	0.5
TEXT_NORMALIZE	True

Dev set results (summary)

Metric	Value
f1_macro_dev_0.5	0.7979572611433388
f1_weighted_dev_0.5	0.8594012496247203
accuracy_dev_0.5	0.8466257668711656
f1_macro_dev_best_global	0.9114611624117328
f1_weighted_dev_best_global	0.9451185172102385
accuracy_dev_best_global	0.9447852760736196
f1_macro_dev_best_by_lang	0.9114611624117328
f1_weighted_dev_best_by_lang	0.9451185172102385
accuracy_dev_best_by_lang	0.9447852760736196
default_threshold	0.5
best_threshold_global	0.7000000000000001
thresholds_by_lang	{"it": 0.7000000000000001}

Thresholds

• Default: 0.5
• Best global: 0.7000000000000001
• Best by language: { "it": 0.7000000000000001 }

Detailed evaluation

Classification report @ 0.5

              precision    recall  f1-score   support

 no-recl (0)     0.9820    0.8258    0.8971       132
    recl (1)     0.5577    0.9355    0.6988        31

    accuracy                         0.8466       163
   macro avg     0.7698    0.8806    0.7980       163
weighted avg     0.9013    0.8466    0.8594       163

Classification report @ best global threshold (t=0.70)

              precision    recall  f1-score   support

 no-recl (0)     0.9695    0.9621    0.9658       132
    recl (1)     0.8438    0.8710    0.8571        31

    accuracy                         0.9448       163
   macro avg     0.9066    0.9165    0.9115       163
weighted avg     0.9456    0.9448    0.9451       163

Classification report @ best per-language thresholds

              precision    recall  f1-score   support

 no-recl (0)     0.9695    0.9621    0.9658       132
    recl (1)     0.8438    0.8710    0.8571        31

    accuracy                         0.9448       163
   macro avg     0.9066    0.9165    0.9115       163
weighted avg     0.9456    0.9448    0.9451       163

Per-language metrics (at best-by-lang)

lang	n	acc	f1_macro	f1_weighted	prec_macro	rec_macro	prec_weighted	rec_weighted
it	163	0.9448	0.9115	0.9451	0.9066	0.9165	0.9456	0.9448

Data

• Train/Dev: private multilingual splits with ~15% stratified Dev (by (lang,label)).
• Source: merged EN/IT/ES data with bios retained (ignored if unused by model).

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoConfig
import torch, numpy as np

repo = "SimoneAstarita/it-no-bio-20251014-t21"
tok = AutoTokenizer.from_pretrained(repo)
cfg = AutoConfig.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo)

texts = ["example text ..."]
langs = ["en"]

mode = "best_global"  # or "0.5", "by_lang"

enc = tok(texts, truncation=True, padding=True, max_length=256, return_tensors="pt")
with torch.no_grad():
    logits = model(**enc).logits
probs = torch.softmax(logits, dim=-1)[:, 1].cpu().numpy()

if mode == "0.5":
    th = 0.5
    preds = (probs >= th).astype(int)
elif mode == "best_global":
    th = getattr(cfg, "best_threshold_global", 0.5)
    preds = (probs >= th).astype(int)
elif mode == "by_lang":
    th_by_lang = getattr(cfg, "thresholds_by_lang", {})
    preds = np.zeros_like(probs, dtype=int)
    for lg in np.unique(langs):
        t = th_by_lang.get(lg, getattr(cfg, "best_threshold_global", 0.5))
        preds[np.array(langs) == lg] = (probs[np.array(langs) == lg] >= t).astype(int)
print(list(zip(texts, preds, probs)))

Additional files

reports.json: all metrics (macro/weighted/accuracy) for @0.5, @best_global, and @best_by_lang. config.json: stores thresholds: default_threshold, best_threshold_global, thresholds_by_lang. postprocessing.json: duplicate threshold info for external tools.