Khalyie/sst2-distilbert

SST-2 Sentiment Classifier (DistilBERT, fine-tuned)

distilbert-base-uncased fine-tuned on GLUE SST-2 (full ~67k train set), 3 epochs, lr 2e-5.

Validation metrics

{
  "eval_loss": 0.5542670488357544,
  "eval_accuracy": 0.9002293577981652,
  "eval_f1": 0.9027932960893855,
  "eval_runtime": 0.9298,
  "eval_samples_per_second": 937.86,
  "eval_steps_per_second": 15.057,
  "epoch": 3.0
}

Files

• config.json, pytorch_model.bin / model.safetensors — full model config and weights (all layers including biases)
• tokenizer files (vocab.txt, tokenizer_config.json, etc.)
• metrics.json — validation accuracy/F1 from training

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
tok = AutoTokenizer.from_pretrained("Khalyie/sst2-distilbert")
model = AutoModelForSequenceClassification.from_pretrained("Khalyie/sst2-distilbert")

See predict_and_visualize_sst2.py in the companion dataset repo (Khalyie/sst2-distilbert-data) for a ready-to-run evaluation + visualization script.

Note: GLUE's official SST-2 test split ships with labels withheld (label == -1); reported metrics above are on the labeled validation split.