README.md

metadata

language:
  - en
  - tr
license: bigcode-openrail-m
tags:
  - stm32
  - embedded
  - code-generation
  - hal
  - firmware
base_model: bigcode/starcoder2-3b
datasets:
  - custom-stm32-dataset
metrics:
  - accuracy
model-index:
  - name: starcoder2-stm32-turkish
    results:
      - task:
          type: text-generation
        metrics:
          - type: accuracy
            value: 85
            name: Test Accuracy (20 realistic projects)

STM32 Code Generator - Turkish Fine-tuned StarCoder2-3B

Model Özeti

Bu model, bigcode/starcoder2-3b temel alınarak 55,000 STM32 kod örneği ile fine-tune edilmiştir. STM32F4 serisi için HAL kütüphanesi kullanan gömülü sistem kodu üretir.

Performans

• Test Sonucu: 17/20 (%85) başarı oranı
• V1 Model: 6/20 (%30) → V2 Model: 17/20 (%85)
• İyileşme: +%55

Test Senaryoları:

• ✅ ADC okuma ve işleme
• ✅ PWM motor kontrolü
• ✅ UART iletişim
• ✅ I2C/SPI protokolleri
• ✅ Timer interrupt'lar
• ✅ DMA kullanımı
• ✅ Watchdog timer
• ✅ Multi-sensor sistemler

Kullanım

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

print("🔄 Model yükleniyor...")

base_model = AutoModelForCausalLM.from_pretrained(
    "bigcode/starcoder2-3b",
    torch_dtype=torch.float16,  
    device_map="auto"
)

model = PeftModel.from_pretrained(base_model, "MuratKomurcu/starcoder2-stm32-turkish")
tokenizer = AutoTokenizer.from_pretrained("MuratKomurcu/starcoder2-stm32-turkish")

print("✅ Model hazır!")

# TEST
prompt = '''### Instruction:
Write complete STM32F401RE LED blink code with main function

### Input:
LED on PA5, blink every 1000ms, include SystemClock_Config

### Response:
'''

Eğitim Detayları

• Base Model: bigcode/starcoder2-3b
• Dataset Size: 55,000 STM32 örnekleri
• Method: LoRA fine-tuning (QLoRA 4-bit)
• LoRA Rank: 64
• Training Epochs: 5
• Final Validation Loss: 0.0297
• Hardware: NVIDIA A100 (40GB)
• Training Time: ~36 saat

Hyperparameters:

• Learning rate: 2e-4
• Batch size: 8 (effective: 32 with gradient accumulation)
• Weight decay: 0.01
• LR Scheduler: Cosine
• Warmup ratio: 0.1

Limitasyonlar

• STM32F4 serisine odaklıdır
• HAL kütüphanesi kullanır (LL veya register-level değil)
• Bazı kompleks senaryolarda (RGB multi-channel PWM, dual button) daha az başarılı
• Türkçe açıklamalar kabul eder ancak İngilizce promptlar daha iyi sonuç verir

Lisans

BigCode OpenRAIL-M

Atıf

@misc{starcoder2-stm32-turkish,
  author = {Murat},
  title = {STM32 Code Generator - Turkish Fine-tuned},
  year = {2024},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/MuratKomurcu/starcoder2-stm32-turkish}}
}