AmanC007/Spiking-CODER-V2

Spiking-CODER v2 🧠⚡

Spiking-CODER v2 is an improved fine-tuned LLM for SNN programming, trained on a hybrid dataset that emphasizes human-written code.

---

📊 Training Data

• Hybrid-v2 Dataset:
  • 40K raw SNN code samples from GitHub.
  • Dataset heavily features Brian2 code, with additional coverage of snnTorch and Norse.
  • Prompts: Synthetic, generated by Mistral-Large.
  • Outputs: Always human-written SNN code.

---

🎯 Key Features

• More realistic training distribution by keeping outputs strictly human-authored.
• Training objective focused on minimizing runtime errors and improving functional pass rates.
• Better alignment with runnable examples across SNN Frameworks (Brian2, snnTorch, Norse, etc.).
• Higher composite evaluation score and Functional Pass rates on SNNBench compared to v1.

---