🦉 Sensible — ModernBERT Sentiment Analysis

The modern replacement for the classic SST-2 sentiment model — 0.946 vs 0.913 on the exact same benchmark, one `pipeline()` line.

from transformers import pipeline

clf = pipeline("text-classification", model="AnkitAI/Sensible-ModernBERT-Sentiment-Analysis")
clf("This movie was absolutely wonderful!")
# [{'label': 'positive', 'score': 0.99}]

positive / negative for reviews, comments, feedback, social text. Built on ModernBERT-base — Flash-Attention-fast, 149M params, CPU-friendly.

Benchmarks

SST-2 official validation set (872 examples) — the same split every SST-2 model reports on:

Model	Accuracy
💬 This model	0.9461
distilbert-base-uncased-finetuned-sst-2-english (the 3.9M-downloads/month default)	0.9130

+3.3 points over the model most pipelines still default to — from an encoder released five years later. Training script and raw eval outputs ship in this repo; the reported split was never used for training or checkpoint selection.

Labels

id	label
0	negative
1	positive

Batch scoring:

texts = ["Best purchase I've made all year.",
         "Waited 40 minutes and the order was still wrong."]
for t, r in zip(texts, clf(texts, batch_size=64)):
    print(f"{r['label']:<9} {r['score']:.2f}  {t}")

Built for

Product & review analytics — score feedback streams at scale
Social/comment moderation dashboards — fast, CPU-deployable
Drop-in upgrade — same task and label semantics as the distilbert-sst2 default your stack probably uses

Good to know

Two classes only (no neutral) — SST-2 convention; genuinely neutral text gets forced to a side
English, sentence/short-paragraph level
Trained on movie-review sentences (SST-2); transfers well to general reviews/comments, less so to domain jargon — for financial text use FinSense

Training details

Full fine-tune of ModernBERT-base on SST-2 (GLUE, 67k sentences): 2 epochs, lr 2e-5, batch 32, fp32, best checkpoint by held-back 5% of train — the official validation set stayed untouched until final reporting.

Support the Project

If this model is useful in your work, you can support independent research:

Citation

@misc{sensiblesentiment2026,
  author = {Aglawe, Ankit},
  title = {Sensible: ModernBERT Sentiment Analysis},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/AnkitAI/Sensible-ModernBERT-Sentiment-Analysis}
}

Base & license

Apache-2.0 (ModernBERT-base, Answer.AI). Trained on SST-2 (Socher et al., 2013 / GLUE).

More from AnkitAI

Model	Task	Score
FinSense ModernBERT	financial news sentiment (3-class)	0.8675
FinSense distilbert v2	financial news sentiment, tiny	0.8447
Parable	local agent LLMs (GGUF)	—

Version history

v1 (2026-07-20) — initial release: ModernBERT-base, SST-2, seed 42.

More on the Sensible models: ankitaglawe.com/sensible

AnkitAI/Sensible-ModernBERT-Sentiment-Analysis