NOKHAB-Lab/LLM_4_HW-SW_Interface

---
license: cc-by-4.0
task_categories:
- text-generation
language:
- en
tags:
- code
- embedded-systems
- raspberry-pi
- hardware-software-interface
- c-programming
- fine-tuning
- llm
pretty_name: HW/SW Interface LLM Dataset
size_categories:
- 10K<n<100K
---

# HW/SW Interface LLM Dataset

## Overview

A dataset of **~17,000 compiler-validated C programs** for ARM-based embedded platforms (Raspberry Pi 3B+, Zero W, 4B, 5), designed for fine-tuning LLMs on hardware-software interface generation tasks.

Fine-tuned open-source models (6–7B parameters) trained on this dataset achieve **81–90% accuracy**, matching or exceeding GPT-4o (89.3%) and Gemini Flash 2.0 (87.2%).

## Dataset

`data/training_set.jsonl` — **~17,000 instruction-response pairs**, fine-tuning ready.

Combines two sources:
- **~16,000** compiler-validated synthetic programs (Gemini Flash 2.0 + GCC-in-the-loop)
- **~1,000** real-world programs collected from GitHub and educational resources

Each entry:
```json
{"prompt": "Write a C program for Raspberry Pi 4B to read a DHT22 sensor...",
"completion": "#include <wiringPi.h>\n..."}
```

| Category | Count | % |
|---|---|---|
| Sensor integration | 8,559 | 50.5% |
| Combined sensor-actuator | 5,582 | 33.0% |
| Actuator control | 1,009 | 6.0% |
| Real-world | 1,053 | 6.2% |
| Other | 736 | 4.3% |

Covers **400+ distinct sensor types** across diverse communication protocols (SPI, I²C, UART, 1-Wire) and Raspberry Pi hardware models.

## Replication Package

Full replication package (pipelines, fine-tuning notebooks, evaluation scripts, results):
👉 [github.com/NOKHAB-Lab/LLM_4_HW-SW_Interface](https://github.com/NOKHAB-Lab/LLM_4_HW-SW_Interface)

## License

[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)