This model is a fine-tuned version of microsoft/phi-1_5 on zelalt/scientific-papers-3.5-withprompt dataset. It achieves the following results on the evaluation set:

• Loss: 2.1587

Requirements

!pip install accelerate transformers einops datasets peft bitsandbytes

Test Dataset

If you prefer, you can use test dataset from zelalt/scientific-papers or zelalt/arxiv-papers or read your pdf as text with PyPDF2.PdfReader then give this text to LLM with adding "What is the title of this paper?" prompt.

from datasets import load_dataset

test_dataset = load_dataset("zelalt/scientific-papers", split='train')
test_dataset = test_dataset.rename_column('full_text', 'text')

def formatting(example):
    text = f"What is the title of this paper? {example['text'][:180]}\n\nAnswer: "
    return {'text': text}

formatted_dataset = test_dataset.map(formatting)

Sample Code

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

peft_model_id = "zelalt/titletor-phi_1-5"
config = PeftConfig.from_pretrained(peft_model_id)
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path,trust_remote_code=True)
model = PeftModel.from_pretrained(model, peft_model_id)

#from dataset
inputs = tokenizer(f'''{formatted_dataset['text'][120]}''', return_tensors="pt", return_attention_mask=False)
outputs = model.generate(**inputs,max_new_tokens=50, pad_token_id = tokenizer.eos_token_id, eos_token_id = tokenizer.eos_token_id)
text = tokenizer.batch_decode(outputs)[0]
print(text)

#as string
inputs = tokenizer(f'''What is the title of this paper? ...[your pdf as text]..\n\nAnswer: ''', return_tensors="pt", return_attention_mask=False)
outputs = model.generate(**inputs,max_new_tokens=50, pad_token_id = tokenizer.eos_token_id, eos_token_id = tokenizer.eos_token_id)
text = tokenizer.batch_decode(outputs)[0]
print(text)

Notes

• After running it for the first time and loading the model and tokenizer, you can only run generating part to avoid RAM crash.

Output

Input:

What is the title of this paper? Bursting Dynamics of the 3D Euler Equations\nin Cylindrical Domains\nFrançois Golse ∗ †\nEcole Polytechnique, CMLS\n91128 Palaiseau Cedex, France\nAlex Mahalov ‡and Basil Nicolaenko §\n\nAnswer:

Output from LLM:

What is the title of this paper? Bursting Dynamics of the 3D Euler Equations
in Cylindrical Domains
François Golse ∗ †
Ecole Polytechnique, CMLS
91128 Palaiseau Cedex, France
Alex Mahalov ‡and Basil Nicolaenko §

Answer:  Bursting Dynamics of the 3D Euler Equations in Cylindrical Domains<|endoftext|>

Training and evaluation data

Train and validation dataset: zelalt/scientific-papers-3.5-withprompt

Training procedure

Training hyperparameters

• total_train_batch_size: 8
• lr_scheduler_type: cosine

Framework versions

• Transformers 4.35.2
• Pytorch 2.1.0+cu118
• Datasets 2.15.0
• Tokenizers 0.15.0