distilbert-sst2-sentiment (mirror)

DistilBERT fine-tuned on SST-2 for binary sentiment classification (NEGATIVE / POSITIVE). Reports ~91% accuracy on the SST-2 dev set at roughly half the size and twice the speed of BERT-base — the standard default for quick English sentiment work.

This is a mirror. The weights and tokenizer files here are an unmodified copy of distilbert/distilbert-base-uncased-finetuned-sst-2-english, re-hosted on this profile for reproducibility and convenience. All credit for the original work belongs to its authors. The upstream license (apache-2.0) is preserved and applies to this copy. If you need the canonical version, please use the upstream repository.

This mirror carries only the safetensors weights and tokenizer (~268 MB); the upstream ONNX, TensorFlow and Rust variants were left out to keep the repo small.

Usage

from transformers import pipeline

clf = pipeline("sentiment-analysis", model="priyaganesh2050/distilbert-sst2-sentiment")
print(clf("A gorgeous, witty film that earns every one of its two hours."))
# [{'label': 'POSITIVE', 'score': 0.9998}]

Batched, with explicit control:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tok = AutoTokenizer.from_pretrained("priyaganesh2050/distilbert-sst2-sentiment")
model = AutoModelForSequenceClassification.from_pretrained("priyaganesh2050/distilbert-sst2-sentiment")

texts = ["Loved every minute.", "Painfully slow and predictable."]
batch = tok(texts, padding=True, truncation=True, return_tensors="pt")
with torch.no_grad():
    probs = model(**batch).logits.softmax(-1)

for t, p in zip(texts, probs):
    print(f"{t:35s} -> {model.config.id2label[int(p.argmax())]} ({p.max():.3f})")

Pairs directly with priyaganesh2050/rotten-tomatoes-sentiment for zero-shot evaluation, since both use the same movie-review, binary-sentiment setup.

Limitations

• Binary only — there is no neutral class. Neutral or mixed text is forced to one side, usually with a confident-looking score. Do not read the softmax as calibrated confidence.
• Domain — trained on movie-review sentences. Accuracy drops on tweets, support tickets, clinical or financial text.
• Bias — the upstream card documents biased associations (for example, sentiment shifting with the country named in an otherwise neutral sentence). Audit before any consequential use.
• Sentence-length input works best; long documents should be split and aggregated.