Warn about config mismatch for pre-training (#2)

- Warn about config mismatch for pre-training (d451554b67f6166d87fbc24e531ae0bafbdae139)
- Update README.md (72aaf32cc7c56230df94f899d9e66bf3880f502c)
- Update README.md (a050549b8d5f3e26eed43efc46012ee7c5e6d8f7)


Co-authored-by: Nathan Godey <nthngdy@users.noreply.huggingface.co>

README.md

@@ -5,6 +5,8 @@ thumbnail: https://huggingface.co/front/thumbnails/google.png
 license: apache-2.0
 ---
 
+**WARNING**: This is the official generator checkpoint as in the [ELECTRA original codebase](https://github.com/google-research/electra). However, this model is not scaled properly for pre-training with [google/electra-small-discriminator](https://huggingface.co/google/electra-small-discriminator). The paper recommends a hyperparameter multiplier of 1/4 between the discriminator and generator for this given model to avoid training instabilities. This would not be the case when using `google/electra-small-generator` and `google/electra-small-discriminator`, which are similar in size.
+
 ## ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators
 
 **ELECTRA** is a new method for self-supervised language representation learning. It can be used to pre-train transformer networks using relatively little compute. ELECTRA models are trained to distinguish "real" input tokens vs "fake" input tokens generated by another neural network, similar to the discriminator of a [GAN](https://arxiv.org/pdf/1406.2661.pdf). At small scale, ELECTRA achieves strong results even when trained on a single GPU. At large scale, ELECTRA achieves state-of-the-art results on the [SQuAD 2.0](https://rajpurkar.github.io/SQuAD-explorer/) dataset.