📝 Supra-Title-Flan-85M

Supra-Title-Flan-85M is a compact 85M-parameter seq2seq model built by SupraLabs, fine-tuned for automatic title generation from text passages. Created for product deploying really soon, Built on a T5-style encoder-decoder architecture and instruction-tuned in the Flan style, it takes a passage of text and produces a concise, relevant title.

This model is part of the SupraLabs open-source ecosystem alongside the Supra-50M causal series, but takes a different route: instead of next-token text generation, it focuses on structured text-to-text transformation.

---

🏗️ Architecture

Parameter	Value
Architecture	T5ForConditionalGeneration
Parameters	~85M
d_model	512
d_ff	1024
d_kv	64
Encoder layers	8
Decoder layers	8
Attention heads	6
Max sequence length	512
Vocabulary size	32,128
Activation	Gated-GELU
Tokenizer	T5Tokenizer
Precision	float32

---

🚀 Inference

import torch
from transformers import T5ForConditionalGeneration, T5Tokenizer

MODEL_ID = "SupraLabs/Supra-Title-Flan-85M"

tokenizer = T5Tokenizer.from_pretrained(MODEL_ID)
model = T5ForConditionalGeneration.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float32,
)
model.eval()

def generate_title(text: str, max_new_tokens: int = 32, num_beams: int = 4) -> str:
    prompt = f"generate title: {text.strip()}"
    inputs = tokenizer(
        prompt,
        return_tensors="pt",
        max_length=512,
        truncation=True,
    )
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            num_beams=num_beams,
            early_stopping=True,
        )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)


text = """
The James Webb Space Telescope captured its deepest infrared image of the universe,
revealing thousands of galaxies, some dating back to just a few hundred million years
after the Big Bang.
"""

print(generate_title(text))

---

💬 Sample Outputs

Samples:

============================================================
             GENERATED CHAT TITLES
============================================================

User-Prompt 1: Can you give me a recipe for a chocolate cake that doesn't use eggs?
Chat-Title 1 : Chocolate Cake Recipe
------------------------------------------------------------
User-Prompt 2: Explain the difference between a linear regression and a logistic regression model simply.
Chat-Title 2 : Linear regression and logistic regression
------------------------------------------------------------
User-Prompt 3: Write a python script that connects to an API and downloads JSON data.
Chat-Title 3 : Python Script for JSON Data
------------------------------------------------------------
User-Prompt 4: What are the top 5 tourist attractions in Tokyo, Japan?
Chat-Title 4 : Tokyo Tourist Locations
------------------------------------------------------------
User-Prompt 5: Translate the following sentence into Spanish: 'Where is the nearest train station?'
Chat-Title 5 : Translation To Spanish
------------------------------------------------------------
User-Prompt 6: How does photosynthesis work in plants? Explain step by step.
Chat-Title 6 : Photosynthesis in Plants
------------------------------------------------------------
User-Prompt 7: Help me write a professional email asking for a deadline extension on my project.
Chat-Title 7 : Project Date Extension Email
------------------------------------------------------------
User-Prompt 8: What is the capital of Australia and some quick facts about it?
Chat-Title 8 : Australia Capital Facts
------------------------------------------------------------
User-Prompt 9: Solve this math problem: if 3x + 5 = 20, what is the value of x?
Chat-Title 9 : Equation of x
------------------------------------------------------------
User-Prompt 10: Give me a short summary of the main plot of Shakespeare's Hamlet.
Chat-Title 10 : Shakespeare's Hamlet Summary
------------------------------------------------------------
User-Prompt 11: Bruh, my wifi keeps disconnecting the whole time last week 😭
Chat-Title 11 : Wireless Disconnecting Time
------------------------------------------------------------

---

⚠️ Limitations

• Output quality degrades significantly on very short inputs (fewer than ~30 tokens).
• The model may produce generic or repetitive titles for highly specialized topics outside its training distribution.
• Not designed for multilingual input; optimized for English text.

---

📄 License

Released under the Apache 2.0 License.

---

huggingface.co/SupraLabs