|
|
--- |
|
|
license: mit |
|
|
library_name: transformers, llama-cpp-python |
|
|
tags: |
|
|
- dual-model |
|
|
- edge-ai |
|
|
- instruction-tuned |
|
|
- robotics |
|
|
- quantized |
|
|
- raspberry-pi |
|
|
- llama |
|
|
- intent-classification |
|
|
- text-generation |
|
|
- on-device |
|
|
language: |
|
|
- en |
|
|
- ta |
|
|
- hi |
|
|
datasets: |
|
|
- Apex-X/PRODIGY-LAB_SARA |
|
|
--- |
|
|
 |
|
|
 |
|
|
 |
|
|
 |
|
|
 |
|
|
 |
|
|
|
|
|
<p align="center"> |
|
|
<img src="https://huggingface.co/blog/assets/171_prodigy_hf/thumbnail.png" alt="PRODIGY 1.2B Banner" width="80%"> |
|
|
</p> |
|
|
|
|
|
|
|
|
# Model Card for PRODIGY-DOIECHI & PRODIGY-SARA |
|
|
|
|
|
A revolutionary **dual-model AI system** optimized for **edge devices** (Raspberry Pi, Jetson Nano, etc.), combining an ultra-lightweight intent classifier (**PRODIGY-DOIECHI**) with a powerful reasoning engine (**PRODIGY-SARA**). The system intelligently routes queries to the optimal model based on complexity, enabling both **sub-100ms responses** and **deep reasoning** on low-resource hardware. |
|
|
|
|
|
- **GitHub**: [https://github.com/Apex-X/PRODIGY-DOIECHI-SARA](https://github.com/Apex-X/PRODIGY-DOIECHI-SARA) |
|
|
- **Author**: Aadhithya (Apex-X) |
|
|
- **Contact**: [email protected] |
|
|
- **License**: MIT |
|
|
|
|
|
--- |
|
|
|
|
|
## Model Details |
|
|
|
|
|
### 🚀 PRODIGY-DOIECHI |
|
|
- **Type**: Neural Network Intent Classifier |
|
|
- **Parameters**: 1.2M |
|
|
- **Format**: PyTorch (`.pth`) |
|
|
- **Size**: < 50 MB |
|
|
- **Input**: Natural language query |
|
|
- **Output**: Structured intent + confidence score |
|
|
- **Use Case**: Commands, calculations, system operations, greetings |
|
|
|
|
|
### 🧠 PRODIGY-SARA |
|
|
- **Base Model**: Llama-7B ,QWEN-238B |
|
|
- **Quantization**: GGUF `Q4_K_M` |
|
|
- **Format**: GGUF (via `llama-cpp-python`) |
|
|
- **Size**: ~3.8 GB |
|
|
- **Context Window**: 1024 tokens |
|
|
- **Fine-tuned on**: [Apex-X/PRODIGY-LAB_SARA](https://huggingface.co/datasets/Apex-X/PRODIGY-LAB_SARA) |
|
|
- **Domains**: Robotics, Tamil Nadu culture, agriculture, medical support, ethics, general reasoning |
|
|
|
|
|
--- |
|
|
|
|
|
## Intended Use |
|
|
|
|
|
This system is designed for **on-device AI assistants** in: |
|
|
- Smart homes |
|
|
- Educational robots |
|
|
- Industrial IoT monitoring |
|
|
- Rural digital kiosks (supporting Tamil, Hindi, English) |
|
|
|
|
|
**Not intended** for high-stakes medical diagnosis, legal advice, or autonomous weapon systems. |
|
|
|
|
|
--- |
|
|
|
|
|
## Performance (Raspberry Pi 4) |
|
|
|
|
|
| Metric | DOIECHI | SARA | |
|
|
|----------------------|-------------|--------------| |
|
|
| RAM Usage | 45 MB | 3.8 GB | |
|
|
| Avg. Latency | 85 ms | 2.3 sec | |
|
|
| Throughput | 11.8 q/s | 0.43 q/s | |
|
|
| Intent Accuracy | 89% | — | |
|
|
| Generation Speed | — | 2.3 tok/sec | |
|
|
|
|
|
> Combined system averages **0.8s response time** in real-world mixed workloads. |
|
|
|
|
|
--- |
|
|
|
|
|
## Smart Routing Logic |
|
|
|
|
|
The system uses a 4-stage pipeline: |
|
|
1. **Memory** → short-term context (non-persistent) |
|
|
2. **Function Executor** → direct system commands |
|
|
3. **DOIECHI** → classify intent & complexity |
|
|
4. **SARA** → activated only if confidence < 0.85 or query contains reasoning keywords (`explain`, `how`, `why`, etc.) |
|
|
|
|
|
Routing decision takes **< 5ms** and learns from usage patterns. |
|
|
|
|
|
--- |
|
|
|
|
|
## How to Use |
|
|
|
|
|
### Python (Combined System) |
|
|
```python |
|
|
from prodigy_system import ProdigyDualSystem |
|
|
system = ProdigyDualSystem() |
|
|
print(system.process("What's 128 / 4?")) # → DOIECHI |
|
|
print(system.process("Explain photosynthesis.")) # → SARA |
|
|
``` |
|
|
--- |
|
|
|
|
|
## 🧠 Hardware Requirements |
|
|
|
|
|
**Minimum** |
|
|
- Raspberry Pi Zero 2 W (64-bit OS) |
|
|
- OS: Linux (64-bit) |
|
|
- Python: 3.8+ |
|
|
- RAM: 2–4 GB |
|
|
- Dependencies: |
|
|
`torch`, `llama-cpp-python`, `nltk`, `psutil` |
|
|
|
|
|
**Recommended** |
|
|
- Raspberry Pi 4 (4GB+ RAM) or NVIDIA Jetson Nano |
|
|
- OS: Ubuntu 20.04+ / Raspberry Pi OS 64-bit |
|
|
- Python: 3.10+ |
|
|
- Dependencies: |
|
|
`torch` (CUDA build), `llama-cpp-python`, `nltk`, `psutil`, `huggingface_hub` |
|
|
|
|
|
**Optional** |
|
|
- Use Hugging Face Spaces or local FastAPI app for deployment |
|
|
- GPU acceleration supported (NVIDIA RTX 20xx or higher) |
|
|
- Convert model to GGUF or quantized formats for faster inference |
|
|
|
|
|
--- |
|
|
|
|
|
## ⚖️ Ethical Considerations |
|
|
|
|
|
- **No Persistent Data Storage** |
|
|
The system does not store personal data or user history beyond the current session. |
|
|
|
|
|
- **User Privacy First** |
|
|
Every interaction is processed locally or in-memory. No external tracking or telemetry. |
|
|
|
|
|
- **Multilingual Accessibility** |
|
|
Built with South Indian language inclusivity in mind, ensuring wider digital access. |
|
|
|
|
|
- **Bias Awareness** |
|
|
Model responses are generated from training data that may contain inherent biases. |
|
|
Always review critical outputs with human oversight. |
|
|
|
|
|
- **Responsible Usage** |
|
|
This model is for research, educational, and robotics-related applications only. |
|
|
Avoid use in contexts that generate harmful, discriminatory, or deceptive content. |
|
|
|
|
|
--- |
|
|
|
|
|
*Developed as part of the **PRODIGY 1.2B** open research initiative on Hugging Face.* |
|
|
*Optimized for lightweight AI deployment on edge devices like Raspberry Pi and Jetson Nano.* |
|
|
|
|
|
Citation |
|
|
@software{prodigy_dual_2025, |
|
|
author = {Aadhithya Ravi}, |
|
|
title = {PRODIGY-DOIECHI \& PRODIGY-SARA: A Dual-Model Edge AI System}, |
|
|
year = {2025}, |
|
|
publisher = {GitHub}, |
|
|
journal = {GitHub repository}, |
|
|
howpublished = {\url{https://github.com/Apex-X/PRODIGY-DOIECHI-SARA}} |
|
|
} |
|
|
--- |
|
|
|
|
|
## 🏁 Acknowledgements |
|
|
|
|
|
- Built on **llama.cpp** and **PyTorch** |
|
|
- Inspired by **Alpaca**, **Self-Instruct**, and **TinyLLM** research |
|
|
- Special thanks to the **Raspberry Pi** and **open-source AI** communities for enabling lightweight, accessible edge AI innovation |
|
|
|
|
|
© 2025 **Aadhithya (Apex-X)**. Released under the **MIT License**. |
|
|
|
|
|
--- |
|
|
|
|
|
This format follows Hugging Face’s **standard model card structure**, includes all metadata in the YAML frontmatter, and is ready to be used as the `README.md` in a Hugging Face **model repository** (e.g., `Apex-X/PRODIGY-DOIECHI-SARA`). |
|
|
|
|
|
Let me know if you'd like separate cards for each model or a version optimized for the **Hugging Face Spaces** demo! |
