Bachstelze/instructionBERTtest

Model Card for the test-version of instructionBERT for Bertology

A minimalistic instruction model with an already good analysed and pretrained encoder like BERT. So we can research the Bertology with instruction-tuned models, look at the attention and investigate what happens to BERT embeddings during fine-tuning.

The trainings code is released at the instructionBERT repository. We used the Huggingface API for warm-starting BertGeneration with Encoder-Decoder-Models for this purpose.

Run the model with a longer output

from transformers import AutoTokenizer, EncoderDecoderModel

# load the fine-tuned seq2seq model and corresponding tokenizer
model_name = "Bachstelze/instructionBERTtest"
model = EncoderDecoderModel.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

input = "Write a poem about love, peace and pancake."
input_ids = tokenizer(input, return_tensors="pt").input_ids
output_ids = model.generate(input_ids, max_new_tokens=200)
print(tokenizer.decode(output_ids[0]))

Training parameters

• base model: "bert-base-cased"
• test subset of the Muennighoff/flan dataset
• trained for 0.97 epochs
• batch size of 14
• 10000 warm-up steps
• learning rate of 0.00005

Purpose of instructionBERT

InstructionBERT is intended for research purposes. The model-generated text should be treated as a starting point rather than a definitive solution for potential use cases. Users should be cautious when employing these models in their applications.