cat_v5c_s2

DeBERTa-v3-base fine-tuned specifically for Tunisian hotel reviews. It tags a review sentence with every category present, using seven independent sigmoid outputs.

Labels, in output order:

Room & Property, Food & Drink, Staff & Service, Location, Pool & Beach, Spa & Wellness, Entertainment & Animation.

Location is kept for label order but is non-actionable: the dashboard excludes it from aggregation, the UI and the score.

Training

Hand-labelled public reviews from Tunisian hotels. The checkpoint was scored on a held-out Tunisian property that was absent from training. The training corpus and the evaluation gold are not released.

Results

Held-out Tunisian property, six actionable categories, Location excluded:

micro F1	macro F1
0.903	0.905

Limitations

• English only.
• Hotel domain only; trained on Tunisian hotel reviews.
• Sentence-level: one sentence at a time, and a sentence can carry several categories.
• Maximum input length is 128 tokens; longer text is truncated.
• Location is predicted but non-actionable and excluded from the score.

Usage

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "YoussefHarrabi/cat_v5c_s2"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id).eval()

text = "The room was clean but the air conditioning was noisy."
with torch.no_grad():
    logits = model(**tok(text, return_tensors="pt", truncation=True, max_length=128)).logits
probs = torch.sigmoid(logits)[0]
for label, p in zip(model.config.id2label.values(), probs):
    print(f"{label}: {p:.3f}")

thresholds.json holds the per-label decision thresholds used by the dashboard. from_pretrained does not apply them; compare each sigmoid with its own threshold.

Intended use

Sentence-level category tagging inside an aspect-based review dashboard. It is a multi-label classifier, not a sentiment model and not a whole-review scorer.