microsoft/deberta-v3-small, fine-tuned for emotion detection
This is not microsoft/deberta-v3-small, which is a general encoder with no classification head. It is that model fine-tuned on dair-ai/emotion, and it is the small half of a pair: this one is built for throughput, the DeBERTa-xlarge version for accuracy.
Sensible
AnkitAI/deberta-v3-small-base-emotions-classifierSix-way emotion classification on the dair-ai/emotion dataset. The model emits sadness, joy, love, anger, fear and surprise — that label set is taken from the model's own config.json, which is what it actually returns at inference. The card lists disgust in place of love; the card is wrong on that point and the config is authoritative. Those six are the whole space the head can return: the choice is forced, so text carrying no emotion at all still comes back with a label, and dair-ai/emotion is short English social-media-register text, which puts long documents and formal prose outside the training distribution.
This model exists as the fast half of a deliberate pair. It and the DeBERTa-xlarge version are the same task on the same dataset with the same labels, and the only real question is whether you are constrained by accuracy or by latency. This is the one for high-throughput batch scoring and CPU-bound deployments.
On benchmarks
This page does not quote an accuracy for this model, and the reason is worth stating plainly. Its card reports 94.6% — but the entire evaluation block on that card, including an evaluation runtime given as 110070.6349 seconds, is byte-identical to the block on the DeBERTa-xlarge card. Two models an order of magnitude apart in size do not produce the same runtime to four decimal places, so at most one of those cards is reporting this checkpoint's numbers. Rather than pick, this page reports none and links both cards. If you need a figure for this model, measure it on your own split.
Run it
from transformers import pipeline
clf = pipeline(
"text-classification",
model="AnkitAI/deberta-v3-small-base-emotions-classifier",
)
clf("I can't believe they pulled it off.")
# labels: sadness, joy, love, anger, fear, surpriseTry others
- microsoft/deberta-xlarge-mnli — six-way emotion classificationThe accuracy half of the emotion pair: its card reports 94.6% on the dair-ai/emotion test split.
- answerdotai/ModernBERT-base — general sentiment analysisA drop-in upgrade for the distilbert-sst2 default: 0.9461 against 0.9130 on the same validation set.
Where to get it
- Base model
- microsoft/deberta-v3-small
- Training data
- dair-ai/emotion