text classification?

[arpitest]

I've trained a small model for hungarian language for 5 days, text generation is working well. Is it possible to use this model for text classification too? (after finetraining on 100k text+class pairs dataset somehow) Where/how can be the classification hidden layer added/connected to this model?

[chrisociepa]

Take a look at this tutorial from HF. I see the following steps to do:

1. fine-tune the model with pairs: text (input) and class (output token)
2. inference:

   with torch.no_grad():
   logits, _ = model(**inputs)
   predicted_class_id = logits.argmax().item()

Another way could be using embedddings (model.embeddings(inputs) produces embbedings) and training a classification NN at top of them (instead of using model.lm_head that producing logits).

You can also convert your model to HF using export_to_hf.py and then fine-tune using their trainers and methods.

[arpitest]

thank you! i didnt notice the generate_embeddings(text) in sample.py earlier... that output can be feed into my classifier NN layers.

also the HF-export is a good idea, already converted the model and was able to load using transformers' from_pretrained(), but i'd like to implement myself...