october-project-first-sweep-20251007-142350-t01

Multilingual XLM-T (EN/IT/ES) binary classifier
Task: LGBTQ+ reclamation vs non-reclamation on social media text.

Trial timestamp (UTC): 2025-10-07 14:23:50

Configuration (trial hyperparameters)

Hyperparameter	Value
LR	1e-05
EPOCHS	3
MAX_LENGTH	256
USE_BIO	True
USE_LANG_TOKEN	True
GATED_BIO	True
FOCAL_LOSS	True
FOCAL_GAMMA	1.5
USE_SAMPLER	True
R_DROP	True
R_KL_ALPHA	1.0

Dev set results

Metric	Value
trial	1
cfg	{'LR': 1e-05, 'EPOCHS': 3.0, 'MAX_LENGTH': 256.0, 'USE_BIO': 1.0, 'USE_LANG_TOKEN': 1.0, 'GATED_BIO': 1.0, 'FOCAL_LOSS': 1.0, 'FOCAL_GAMMA': 1.5, 'USE_SAMPLER': 1.0, 'R_DROP': 1.0, 'R_KL_ALPHA': 1.0}
f1_macro_dev	0.7642181107650974
best_threshold_dev	0.85
precision_macro_dev	0.8431519699812383
recall_macro_dev	0.7226866883116883

Data

• Train/Dev: private dataset (merged train_en.csv, train_it.csv, train_es.csv with 15% stratified Dev).
• The dataset id is intentionally not disclosed in metadata.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
tok = AutoTokenizer.from_pretrained("SimoneAstarita/october-project-first-sweep-20251007-142350-t01")
model = AutoModelForSequenceClassification.from_pretrained("SimoneAstarita/october-project-first-sweep-20251007-142350-t01")