sparrowaisolutions/aras-ember-v1

🔥 Aras-Ember v1

Aras-Ember v1 is a lightweight conversational AI model trained on the Ember instruction dataset and built on top of Falcon RW-1B.

This model is designed for chat, instruction following, and experimentation with compact open-source LLMs.

Model page: https://huggingface.co/sparrowaisolutions/aras-ember-v1

---

[] (https://colab.research.google.com/drive/18UVyG5RWn5YJlyZqyEyP_MzvMrYoplqX?usp=sharing)

🧠 Model Details

Model name: Aras-Ember v1 Base model: tiiuae/falcon-rw-1b Architecture: Falcon decoder transformer Parameters: ~1B

Training method:

• LoRA fine-tuning
• merged weights for standalone model

---

📚 Dataset

Training dataset:

→ https://huggingface.co/datasets/sparrowaisolutions/ember-dataset

Dataset format:

{
  "instruction": "...",
  "response": "..."
}

The dataset contains instruction-response pairs designed for conversational AI training.

---

⚙️ Training Details

Training setup:

• Base model: Falcon RW-1B
• Dataset size used: ~15K examples
• Training epochs: 2
• Method: LoRA fine-tuning
• LoRA merged into final weights

Frameworks used:

• Transformers
• PEFT
• PyTorch
• Hugging Face Datasets

---

🚀 Usage

Install dependencies:

!pip install transformers accelerate safetensors gradio torch --upgrade

Example usage:

# Imports
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
import torch
from threading import Thread
import gradio as gr

# Model ID on Hugging Face
model_id = "sparrowaisolutions/aras-ember-v1"

# Load tokenizer & model
tokenizer = AutoTokenizer.from_pretrained(model_id)
tokenizer.pad_token = tokenizer.eos_token  # required for generation padding

model = AutoModelForCausalLM.from_pretrained(model_id)
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

# Default AI prompt
DEFAULT_PROMPT = """You are Aras-Ember, a helpful AI assistant by Sparrow AI Solutions.
Respond clearly and concisely. Stay in chatbot mode and do NOT generate unrelated articles or dataset text.
"""

def generate_stream(user_input):
    """Generate streamed response from model"""
    prompt = DEFAULT_PROMPT + f"User: {user_input}\nAras-Ember:"

    inputs = tokenizer(prompt, return_tensors="pt").to(device)
    
    # Streamer for live output
    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

    generation_kwargs = dict(
        **inputs,
        streamer=streamer,
        max_new_tokens=150,
        do_sample=True,
        temperature=0.6,
        top_p=0.9,
        repetition_penalty=1.2,
        no_repeat_ngram_size=3,
        pad_token_id=tokenizer.eos_token_id
    )

    # Run generation in separate thread
    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    thread.start()

    # Stream output
    response = ""
    for new_text in streamer:
        print(new_text, end="", flush=True)
        response += new_text
    print()  # newline after completion

    return response

# Gradio chat interface
def chat_fn(user_input):
    return generate_stream(user_input)

iface = gr.Interface(
    fn=chat_fn,
    inputs=gr.Textbox(lines=3, placeholder="Say something to Aras-Ember..."),
    outputs="text",
    title="Aras-Ember Chat",
    description="A lightweight AI by Sparrow AI Solutions."
)

# Launch Gradio in Colab
iface.launch(share=True)  # `share=True` gives a public URL for testing

---

💬 Chat Format

Recommended prompt format:

You are Aras-Ember, a creative AI assistant.

User: <question>
Assistant:

Example:

User: Explain black holes simply.
Assistant:

---

🧪 Intended Use

Aras-Ember is intended for:

• conversational AI
• research experiments
• educational projects
• lightweight chatbot systems

Not intended for critical decision-making systems.

---

⚠️ Limitations

• small parameter model (1B)
• limited reasoning compared to larger LLMs
• may hallucinate facts
• trained on a limited dataset subset

---

🏗 Architecture

Base architecture:

→ Falcon RW-1B by the Technology Innovation Institute

Reference model:

https://huggingface.co/tiiuae/falcon-rw-1b

---

👨‍💻 Authors

Created by:

Sparrow AI Solutions

Hugging Face:

https://huggingface.co/sparrowaisolutions

---

❤️ Acknowledgements

Thanks to:

• Hugging Face
• Falcon model creators
• open-source AI community

---

Research Paper

EMBER Dataset and ARAS-EMBER Models: Open Lightweight AI Systems for Creative and Conversational Language Generation

DOI: https://doi.org/10.6084/m9.figshare.31617994

📜 License

Apache 2.0 License