Lura 1.1

Lura 1.1 is an ultra-lightweight language model developed by SEN-AGI. It is the successor to Lura 1.0, fine-tuned from the Qwen 2.5 500M (0.5B) base model.

Lura 1.1 demonstrates that carefully curated, high-quality synthetic data allows compact models to punch significantly above their weight class.

🚀 What's New in Lura 1.1?

Compared to Lura 1.0, Lura 1.1 brings noticeable improvements across key everyday tasks:

Better Instruction Following: Adheres more reliably to formatting, constraints, and multi-turn instructions.
Improved Elementary Math & Logic: Better step-by-step handling of simple arithmetic and logical puzzles.
Basic Code Generation: Capable of writing small scripts, basic functions, and code explanations.
High Efficiency: Runs smoothly on CPUs, edge devices, smartphones, and low-VRAM setups.

🛠️ Training & Dataset

Base Model: Qwen/Qwen2.5-0.5B
Data Strategy: Rather than using generic, multi-gigabyte open datasets, Lura 1.1 was fine-tuned on 4,000–5,000 custom synthetic samples. These samples were generated and filtered using frontier (SOTA) large language models to maximize density, diversity, and reasoning quality.
Zero-Cost Engineering: Trained entirely on Google Colab's free T4 GPU tier and managed directly from a single mobile device without requiring dedicated local GPUs or desktop computers.

💻 How to Use

You can easily run Lura 1.1 using Hugging Face's transformers library:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "SEN-AGI/Lura-1.1-500m"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto"
)

prompt = "Explain why the sky is blue in three simple sentences."
messages = [
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.9
)

response = tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True)
print(response)

⚠️ Limitations & Biases

Because Lura 1.1 is built on a 500M parameter architecture, it has natural
capacity limits:

  - Hallucinations: It may state incorrect information confidently, especially
    on niche or complex factual queries.
  - Complex Reasoning: It is not intended for advanced mathematics, competitive
    programming, or deep multi-hop logic.
  - Recommended Use Cases: Everyday basic assistance, simple
    formatting/summarization, lightweight offline experimentation, and edge
    device deployment.

👨‍💻 Developed By

  - Creator: SEN-AGI
  - Focus: Pushing the limits of compact language models through high-quality
    synthetic data curation.