Hiratax/t5-news-summarizer

T5-Base News Summarizer (Multi-Style)

This model is a fine-tuned version of google/flan-t5-base trained on samples from CNN/DailyMail and XSum.
It generates news summaries in three styles: Harsh (Concise), Standard, and Detailed.

---

Model Description

• Model Type: Sequence-to-Sequence Transformer (T5)
• Language: English
• Base Model: google/flan-t5-base
• Training Data: ~9k mixed samples from CNN/DailyMail & XSum

Key Features

This model supports a Style Prompt that determines summary length and density:

1. Harsh

   • Very concise
   • Headline-like
   • Trained mostly on XSum

2. Standard

   • Balanced, general-purpose summarization

3. Detailed

   • Longer, more contextual summaries
   • Trained with CNN/DailyMail

---

Usage

from transformers import pipeline

summarizer = pipeline("summarization", model="Hiratax/t5-news-summarizer")

text = """
The James Webb Space Telescope (JWST) has captured a lush landscape of stellar birth. 
The new image shows the Cosmic Cliffs, which are the edge of a giant gaseous cavity within the star-forming region NGC 3324.
"""

# 1. Standard
print(summarizer("summarize standard: " + text))

# 2. Harsh (Headline)
print(summarizer("summarize harsh: " + text))

# 3. Detailed
print(summarizer("summarize detailed: " + text))

---

Recommended Inference Parameters

Style	Min Length	Max Length	Length Penalty	Repetition Penalty	N-Gram Block
Harsh	10	35% of input	1.0	2.0	3
Standard	60	150	2.0	1.5	3
Detailed	50% input	150% of input	1.5	1.2	4

Tip:
"Detailed" style benefits from no_repeat_ngram_size=4 to avoid repeated openings.

---

Training Procedure

Hyperparameters

• Epochs: 8
• Learning Rate: 1e-4
• Batch Size: 4
• Gradient Accumulation: 2
• Weight Decay: 0.01
• Optimizer: AdamW
• Precision: FP16

Data Strategy

• Harsh → XSum (abstractive, short)
• Detailed → CNN/DailyMail (longer, higher detail)
• Safety: Removed cases where summary > article length to reduce hallucinations

---

Limitations

• May occasionally output the typo "occupys" (training noise).
• Max input length: 512 tokens (longer text is truncated).
• Model performance decreases on extremely long or highly technical articles.

---

License

Apache 2.0