BESSTIE sarcasm classifier (DeBERTa-v3-base)

Sarcasm classification for Australian, British and Indian English, trained on the BESSTIE benchmark.

184M parameters, ~25 ms per prediction on CPU.

Results

Macro-F1 on the official validation split, in-dialect, mean over 5 seeds.

The majority-class baseline is shown because it is the number that makes the others interpretable: a model that never predicts the minority class scores about 0.46 on this task, so scores near 0.50 indicate no learning.

Important: this model needs a decision threshold

argmax is the wrong operating point. The positive class is a minority, so the threshold was tuned on validation to maximise macro-F1:

• temperature: 1.1819 (applied to logits before softmax)
• decision threshold: 0.360

Both must be applied together and in that order. A threshold fitted on raw probabilities but applied to temperature-scaled ones labels every input positive.

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

name = "ppokhrel2109/besstie-sarcasm-deberta-v3"
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForSequenceClassification.from_pretrained(name).eval()

text = "oh brilliant, another delayed train"
with torch.no_grad():
    logits = model(**tok(text, return_tensors="pt", truncation=True, max_length=128)).logits
probability = torch.softmax(logits / 1.1819, dim=-1)[0, 1].item()
label = "sarcastic" if probability >= 0.360 else "not sarcastic"

Training

Class-weighted cross-entropy, threshold tuned on a held-out slice of train, five seeds. Sarcasm is evaluated on the Reddit subset only, matching the benchmark's protocol.

Limitations

• Evaluated on the official validation split; the benchmark's test split is withheld.
• en-IN is the weakest variety and has the fewest sarcastic training examples.
• Trained on Reddit comments and Google reviews; other domains are out of distribution.

Code, full results and reproduction steps: https://github.com/Pranav210901/Besstie-improvement-attempt