Text Summarization Model

Model Overview

This is a text summarization model built using a Seq2Seq architecture.
It was trained on the CNN/DailyMail dataset (3.0.0) and is capable of generating concise summaries of news articles or other long-form texts.

Intended Use:

  • Summarizing articles, documents, or reports.
  • Extracting key points from text for quick understanding.

Limitations & Biases:

  • May struggle with extremely long articles or highly technical content.
  • Generated summaries may occasionally miss nuanced details.

Training Details

  • Dataset: CNN/DailyMail (3.0.0 version)
  • Preprocessing: Truncation at 512 tokens for input, summaries capped at 150 tokens.
  • Hyperparameters:
    • Optimizer: AdamW (PyTorch)
    • Learning rate: 2e-5
    • Batch size: 4 (per device)
    • Epochs: 10
  • Evaluation Metrics: ROUGE-1, ROUGE-2, ROUGE-L

Evaluation Results

Metric Score (%)
ROUGE-1 83.3
ROUGE-2 60.0
ROUGE-L 83.3
ROUGE-Lsum 83.3

Example Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("your-username/your-model-name")
model = AutoModelForSeq2SeqLM.from_pretrained("your-username/your-model-name")

text = "The stock market saw a significant drop today due to rising inflation concerns. Investors are cautious ahead of the Federal Reserve's upcoming decision."

inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
summary_ids = model.generate(**inputs, max_length=150, num_beams=4, early_stopping=True)

print(tokenizer.decode(summary_ids[0], skip_special_tokens=True))
Downloads last month
7
Safetensors
Model size
139M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Dataset used to train ssarathi/Text_Summarizer_Using_Transformers