|
|
--- |
|
|
license: apache-2.0 |
|
|
tags: |
|
|
- summarization |
|
|
- custom-model |
|
|
- pegasus |
|
|
- seq2seq |
|
|
- huggingface |
|
|
- transformers |
|
|
library_name: transformers |
|
|
inference: false |
|
|
model-index: |
|
|
- name: Custom Pegasus Summarizer |
|
|
results: [] |
|
|
--- |
|
|
|
|
|
# 🦅 Custom Pegasus Summarizer |
|
|
|
|
|
This model is a **custom-wrapped version** of \[`google/pegasus-xsum`\](https://huggingface.co/google/pegasus-xsum) built for **summarization tasks**. It\'s implemented using Hugging Face\'s \`transformers\` library and wrapped with a custom model class for educational and experimental flexibility. |
|
|
|
|
|
✅ It supports: |
|
|
- Easy fine-tuning and extension \(e.g., adapters, prompt tuning\) |
|
|
- Drop-in replacement for the original model |
|
|
- Hugging Face Hub compatibility |
|
|
- Works with \`AutoTokenizer\` and \`CustomSeq2SeqModel\` |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧠 Model Architecture |
|
|
|
|
|
- **Base**: google/pegasus-xsum |
|
|
- **Wrapper**: CustomSeq2SeqModel \(inherits from PreTrainedModel\) |
|
|
- **Tokenizer**: AutoTokenizer from the same repo |
|
|
- **Configuration**: CustomSeq2SeqConfig \(inherits from PretrainedConfig\) |
|
|
|
|
|
--- |
|
|
|
|
|
## 🧪 Training Details |
|
|
|
|
|
- **Dataset**: xsum \(500-sample subset\) |
|
|
- **Task**: Abstractive Summarization |
|
|
- **Epochs**: 1 |
|
|
- **Batch Size**: 4 |
|
|
- **Learning Rate**: 2e-5 |
|
|
- **Training Framework**: Hugging Face Trainer |
|
|
|
|
|
--- |
|
|
|
|
|
## 💡 Usage Example |
|
|
|
|
|
\`\`\`python |
|
|
from transformers import AutoTokenizer |
|
|
from model import CustomSeq2SeqModel # Your custom wrapper |
|
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained("your-username/custom-pegasus-summarizer") |
|
|
model = CustomSeq2SeqModel.from_pretrained("your-username/custom-pegasus-summarizer") |
|
|
|
|
|
text = "summarize: The Apollo program was a major milestone in space exploration..." |
|
|
inputs = tokenizer(text, return_tensors="pt", truncation=True) |
|
|
summary_ids = model.generate(**inputs, max_length=60) |
|
|
print(tokenizer.decode(summary_ids[0], skip_special_tokens=True)) |
|
|
\`\`\` |
|
|
|
|
|
--- |
|
|
|
|
|
## 🎛 Live Demos |
|
|
|
|
|
You can try this model interactively on Hugging Face Spaces: |
|
|
|
|
|
- Gradio App: https://huggingface.co/spaces/your-username/custom-pegasus-gradio |
|
|
- Streamlit App: https://huggingface.co/spaces/your-username/custom-pegasus-streamlit |
|
|
|
|
|
--- |
|
|
|
|
|
## 📦 Files Included |
|
|
|
|
|
- \`config.json\` – Model configuration \(used by \`from_pretrained\`\) |
|
|
- \`pytorch_model.bin\` – Fine-tuned model weights |
|
|
- \`tokenizer_config.json\` – Tokenizer settings |
|
|
- \`vocab.json\` / \`merges.txt\` – Tokenizer vocab \(depends on tokenizer type\) |
|
|
- \`special_tokens_map.json\` – Special tokens for summarization |
|
|
- \`README.md\` – This model card |
|
|
- \`model.py\` – \(if included\) Your \`CustomSeq2SeqModel\` class |
|
|
|
|
|
--- |
|
|
|
|
|
## 📜 License |
|
|
|
|
|
Apache 2.0 — same license as the original \`pegasus-xsum\`. |
|
|
|
|
|
--- |
|
|
|
|
|
## 🙏 Acknowledgments |
|
|
|
|
|
- Hugging Face for \`transformers\`, \`datasets\`, and \`hub\` |
|
|
- Authors of PEGASUS |
|
|
- Educational/Research communities building custom NLP models |
|
|
|
