yousefg/Academ-0.5

Model Card for Model ID

Academ is a fine-tuned BART model for summarizing academic lectures.

To find out how the model was fine-tuned, you can check the notebook on Kaggle: https://www.kaggle.com/code/yousefr/college-lectures-summarization-bart-unsupervised/

Model Details

Model Description

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

• Developed by: Yousef Gamaleldin
• Model type: Summarization
• Language(s) (NLP): English
• License: MIT
• Finetuned from model [optional]: BART Large CNN

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import BartForConditionalGeneration, AutoTokenizer

model = BartForConditionalGeneration.from_pretrained('yousefg/Academ-0.5')
tokenizer = AutoTokenizer.from_pretrained('facebook/bart-large-cnn')

def get_summary(input_ids, attention_mask, context_length):
    
    summaries = []
    for i in range(0, input_ids.shape[1], context_length):
        
        input_slice = input_ids[:, i:i + context_length] if i + context_length <= input_ids.size(1) else input_ids[:, i:]
        attention_mask_slice = attention_mask[:, i:i + context_length] if i + context_length <= attention_mask.size(1) else attention_mask[:, i:]
        
        summary = model.generate(input_slice, attention_mask = attention_mask_slice, max_new_tokens = 1654, min_new_tokens = 250, do_sample = True, renormalize_logits = True)
        summaries.extend(summary[0].tolist())
        
    summaries = tokenizer.decode(summaries, skip_special_tokens = True)
    
    return summaries
    
batch = tokenizer(texts, truncation = False) # make sure to get the transcript from the lecture

input_ids = torch.tensor(batch['input_ids']).unsqueeze(0).to(device)
attention_mask = torch.tensor(batch['attention_mask']).unsqueeze(0).to(device)

summary = get_summary(input_ids, attention_mask, 1654)
print(summary)

Training Details

The model's training used a custom loss function for getting the model into an optimal length (35% chosen as the optimal length).

Training Hyperparameters

• Training regime: bf16 non-mixed precision
• Learning Rate: 0.001
• Weight Decay: 0.01
• Epochs: 4
• Batch Size: 16

Evaluation

The evaluation is based on ROUGE 1 with a change of discounting padding tokens.

Testing Data

The model's test dataset had 289 lectures, mainly from MIT OpenCourseWare.

Results

The model achieved 96% accuracy for ROUGUE 1 in the test dataset, and 93% in the evaluation dataset.

Summary

Academ is a summarization model trained on 2307 lectures, mainly from MIT OpenCourseWare. The model has a max sequence length of 1654, increasing 630 tokens from the original model.