A 150M-parameter Turkish classifier that receives a sentence and an aspect, then predicts negative, neutral, or positive sentiment toward that aspect.

Results

Our model scores 72.74 +/- 2.10% macro-F1 on the ABSA-TR test set. The released checkpoint scores 72.87%.

The quantized int8 version scores 71.89% on ABSA-TR test set.

Usage

python

from transformers import pipeline

classify = pipeline(
    "text-classification",
    model="ytu-ce-cosmos/modernbert-tr-absa-polarity",
)
classify({"text": "Kargo hızlıydı ama ürün kötüydü.", "text_pair": "ürün"})

Training

We finetune ytu-ce-cosmos/modernbert-tr-base with mean pooling and a three-label head using translated M-ABSA pairs, 11,470 native Turkish clothing-review rows, and 17,186 usable pairs from teacher-labeled Turkish reviews. We use learning rate 3e-5, batch size 32, 15 epochs, linear warmup and decay, weight decay 0.01, label smoothing 0.1, and five seeds.

Standalone binaries are available in the encoderfile repo.

License

Apache-2.0.