To build an automated news summarization tool that extracts and condenses real news articles into short, high-quality summaries using a transformer-based model, with robust scheduling, free-tier compatibility, and a modern web interface for daily news browsing.
FeedFlash addresses information overload by creating a concise, readable digest of important daily news stories. Designed for readers, students, and professionals, it keeps you informed without the need to read entire articles.
- Headline Fetching: Uses NewsAPI to pull 30 real-world articles per session (now upgraded to 30 total: 20 international, 10 Indian, no genre restriction).
- Content Extraction: Employs
newspaper3kfor robust article scraping and cleaning. - Summarization: Utilizes a PEFT/LoRA-finetuned, 4-bit quantized and dequantized Flan-T5-Base model, supporting up to 600 input tokens and 300 output tokens for richer, context-aware summaries (~90 words).
- Storage: Summaries and logs are saved in JSON files (
summaries.json,fetch_log.json) for easy access and integration. - User Interface: Gradio web interface lets users browse headlines, read summaries, and access original articles, with auto-scroll to summary and robust error handling.
- Automation: GitHub Actions scheduler fetches and summarizes news every 4 hours (10 AM–10 PM), syncing results to the Hugging Face Space.
- Free-Tier Compatibility: All scripts and workflows are designed to run on free cloud and notebook resources.
- Automated News Fetching: Scheduled collection of 30 articles (20 international, 10 Indian) using NewsAPI.
- Full Article Extraction: Cleans and parses full article text with
newspaper3k. - Transformer-Based Summarization: Summaries generated with a Flan-T5-Base model, fine-tuned via PEFT/LoRA, quantized to 4-bit for efficient training, then dequantized for fast CPU/ONNX inference.
- JSON Storage: Outputs are saved in
summaries.jsonandfetch_log.jsonfor persistence and downstream use. - Modern Gradio UI: Interactive, mobile-friendly interface for headline selection, summary viewing, and link access, with auto-scroll and robust error handling.
- Scheduled Automation: GitHub Actions workflow runs every 4 hours, updating summaries and logs, and syncing to Hugging Face Space.
- Error Handling: Graceful handling of missing or malformed data, API errors, and extraction/summarization failures.
- Selected and explored a relevant news summarization dataset.
- Fine-tuned a Flan-T5-Base model (PEFT/LoRA, 4-bit quantized, dequantized for inference) on Kaggle.
- Scripted automated article fetching with NewsAPI and robust content extraction with
newspaper3k. - Implemented summarization logic and error handling.
- Saved all outputs in JSON format.
- Built a Gradio UI for user-friendly access to summaries, with mobile enhancements.
- Set up a GitHub Actions scheduler for regular updates and automatic syncing to Hugging Face Spaces.
requestsnewspaper3ktransformersgradiojsonhuggingface-hubnewsapi-pythontorchsentencepiecepandaslxml
Add your own screenshots here:
| Model | Loss | Training Steps | Notes |
|---|---|---|---|
| T5-Base | 1.59 | 9375 | Base model, 412→100 tokens |
| Flan-T5-Base (LoRA, 4-bit) | 1.95 | 14,517 | Quantized 4-bit, dequantized for CPU, 600→300 tokens |
- Practical experience with transformer-based summarization and quantization.
- Integration of external APIs and web scraping for real-world data.
- Construction of an end-to-end ML pipeline with a user-facing UI and scheduled automation.
- Daily news summarization for students, readers, and professionals.
- Compact news feeds for low-data or accessibility-focused environments.
- Deploy the Gradio app on Hugging Face Spaces or a Flask server.
- Schedule the fetch-and-summarize script to run throughout the day for fresh content.
- Optionally add persistent storage and user feedback logging.
- User feedback logging (thumbs-up/thumbs-down)
- Multi-language or regional news support
Arihant Bhandari
