Base model: cardiffnlp/twitter-roberta-base (~125M, RoBERTa pretrained on tweets)
Dataset: cardiffnlp/tweet_eval (sentiment config, 45.6K train)
Training: 3 epochs, lr 2e-5, batch 32, max_len 128, warmup 0.1, weight decay 0.01

Test-set results (TweetEval sentiment, 12,284 tweets)

Metric	Score
Accuracy	0.7155
Macro-F1	0.7155
Macro-Recall	0.7268
Speed (T4)	~1600 tweets/s

Comparison vs distilbert-base

Model	Size	Accuracy	Macro-F1	tweets/s
twitter-roberta-base (this)	125M	0.7155	0.7155	1600
distilbert-base	67M	0.6888	0.6877	2897

Domain pretraining on tweets gives +2.7 pts accuracy / +2.8 pts macro-F1 over generic DistilBERT, at ~1.8× the inference cost.

Usage

python

from transformers import pipeline
clf = pipeline("text-classification", model="Ido-shraga/twitter-roberta-base-tweeteval-sentiment")
clf("I can't believe how good this is 🔥")