What are your experiences for appropriate dataset sizes for usual text classification tasks using a finetuned BERT such as sentiment analysis?

~100 examples

~1000 examples


~10000000 examples

What are your experiences?


2 Answers 2


Sorry, but there’s no rule and amount we are able to quantify.

I’ve used it (multilingual) for 700 texts with a 20 multilabel classification and I had worse results than with a custom deep net (but with pretrained word embeddings).

But you know, in fact these questions are hard to answer. Why? Because with very low quantities of data, you can’t accurately assess the performance of a model - it is statistically insignificant.

  • $\begingroup$ I also think that it really depends on the downstream task. However, i think it might be helpful to get some impressions what works and what not. $\endgroup$
    – fhaase
    Sep 6, 2019 at 9:34
  • $\begingroup$ I am not really talking about the downstream task, but more about evaluating the results. You can't significantly quantify the performance of a model with a very low sample size. $\endgroup$
    – Elliot
    Sep 6, 2019 at 9:47

The performance of the model is largely dependent on the data it is fed and (Suchdev et al., 2014) recommended a minimum of 5600 tweets as training data for a specific domain.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.