MODEL_CARD.md

metadata

language:
  - en
license: apache-2.0
tags:
  - zig
  - code
  - programming
  - lora
  - qwen2.5-coder
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
model_type: qwen2.5
library_name: transformers

ZigNet Qwen2.5-Coder-7B

Fine-tuned Qwen2.5-Coder-7B for Zig programming language analysis and assistance

This model is part of the ZigNet project - an MCP (Model Context Protocol) server that provides intelligent Zig code analysis for Claude and other LLMs.

Model Details

• Base Model: Qwen/Qwen2.5-Coder-7B-Instruct
• Fine-tuning Method: QLoRA (4-bit quantization)
• Training Data: 13,756 Zig code examples from official documentation (v0.13-0.15)
• Supported Zig Versions: 0.13.x, 0.14.x, 0.15.x
• Training Hardware: NVIDIA RTX 3090 (24GB VRAM)
• Adapter Size: ~155MB (LoRA adapters only)

Training Configuration

LoraConfig:
  - r: 16
  - lora_alpha: 32
  - target_modules: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
  - lora_dropout: 0.05
  - bias: "none"

TrainingArguments:
  - num_train_epochs: 3
  - per_device_train_batch_size: 16
  - learning_rate: 2e-4
  - warmup_steps: 100
  - fp16: true

Dataset

The model was trained on a curated dataset of Zig examples including:

• Official Zig documentation examples (v0.13, v0.14, v0.15)
• Advanced features: comptime, generics, error handling, async
• Real-world code patterns from popular Zig projects

Dataset: fulgidus/zignet-training-dataset

Intended Use

This model is designed to:

• 📖 Provide Zig documentation context
• 💡 Suggest intelligent code fixes for Zig errors
• 🔍 Explain Zig-specific idioms and patterns
• ⚡ Generate idiomatic Zig code

Note: This model is NOT used for parsing or validation (handled by deterministic compiler-based tools). It focuses on documentation lookup and intelligent suggestions.

Performance

• Quality: ⭐⭐⭐⭐⭐ Best-in-class for Zig syntax and idioms
• Benchmarks: 100% pass rate on Zig validation tests
• Response Time: ~15-20s (after GGUF quantization)

Usage

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-7B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load LoRA adapters
model = PeftModel.from_pretrained(base_model, "fulgidus/zignet-qwen2.5-coder-7b")

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct")

# Generate
prompt = "Explain Zig comptime feature with an example"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_length=500)
print(tokenizer.decode(outputs[0]))

With ZigNet MCP Server

This model is integrated into ZigNet for use with Claude:

# Install ZigNet
git clone https://github.com/fulgidus/zignet
cd zignet
pnpm install
pnpm run build

# Configure MCP client (Claude Desktop)
# Add to ~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "zignet": {
      "command": "node",
      "args": ["/path/to/zignet/dist/mcp-server.js"]
    }
  }
}

Limitations

• Focused on Zig 0.13-0.15 (may have limited accuracy on very old or very new syntax)
• LoRA adapters only (requires base model for inference)
• Optimized for English documentation and comments
• Not suitable for real-time parsing (use ZigNet's AST parser for that)

Citation

@software{zignet2025,
  author = {fulgidus},
  title = {ZigNet: Intelligent Zig Code Analysis via MCP},
  year = {2025},
  url = {https://github.com/fulgidus/zignet}
}

License

Apache-2.0 (same as base model)

Acknowledgments

• Base Model: Qwen2.5-Coder by Alibaba Cloud
• Zig Language: ziglang.org
• Training Framework: HuggingFace Transformers + PEFT

---

Project: github.com/fulgidus/zignet
Author: fulgidus
Date: October 2025