Hiratax/t5-base-finetuned-summarizer

T5 Multi-Style Text Summarizer

A fine-tuned version of google/flan-t5-base capable of generating summaries in three different styles based on desired length and detail: Harsh, Standard, and Detailed. The model uses special style prefixes to control output length.

---

🚀 Model Description

• Model type: Text-to-Text (T5)
• Language: English
• Base model: google/flan-t5-base
• Task: Controlled text summarization
• Training datasets: CNN/DailyMail (3.0.0) + XSum
• Total samples: 5,000
• Best use case: News-style text summarization

---

🏷️ Style Prefixes

Prepend one of the following to your input text:

Style	Prefix	Description
Harsh	summarize harsh:	Very concise (~30% of original length)
Standard	summarize standard:	Balanced (~50–60%)
Detailed	summarize detailed:	More comprehensive (~80%)

---

📌 How to Use (Transformers)

from transformers import T5Tokenizer, T5ForConditionalGeneration

model_name = "Hiratax/t5-base-finetuned-summarizer"  
tokenizer = T5Tokenizer.from_pretrained(model_name)
model = T5ForConditionalGeneration.from_pretrained(model_name)

input_text = """
Russian President Vladimir Putin said that a US plan to end the war in Ukraine could 
“form the basis for future agreements” but renewed threats to seize more territory...
"""

# Add style prefix
prefix = "summarize standard: "
inputs = tokenizer(prefix + input_text, return_tensors="pt", max_length=512, truncation=True)

outputs = model.generate(
    inputs["input_ids"],
    max_length=150,
    num_beams=4,
    length_penalty=1.0,
    early_stopping=True
)

summary = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(summary)

---

🧪 Training Details

• Epochs: 5
• Batch size: 4 (Gradient Accumulation = 2)
• Learning rate: 3e-4
• Optimizer: AdamW (weight decay 0.01)
• Max input length: 512 tokens
• Max target length: 150 tokens

Dataset Labeling Logic

Style was determined based on:

summary_length / original_length

• Low ratio → Harsh
• Medium ratio → Standard
• High ratio → Detailed

---

⚠️ Limitations

• Optimized for news-style content.
• Inputs longer than 512 tokens are truncated.
• Style control is approximate.
• Not ideal for creative or conversational text.

---

📄 License

Specify your preferred license, e.g.:

Apache 2.0