haowu1234/signal-dsl-generator-lora

Signal DSL Generator (LoRA Adapter)

A fine-tuned LoRA adapter for generating Signal DSL configurations from natural language descriptions.

Model Details

Attribute	Value
Base Model	Qwen/Qwen2.5-Coder-7B-Instruct
Training Method	3-Stage (SFT → Preference Tuning → DPO)
Final Accuracy	96.25%
DPO Margins	1.57
Parameters	LoRA rank=16, alpha=32

What is Signal DSL?

Signal DSL is a domain-specific language for configuring intelligent LLM routing. It allows you to define:

• Signals: Domain, modality, complexity, language detection
• Routes: Conditional model selection based on signals
• Plugins: System prompts, RAG, semantic cache, etc.

Usage

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

# Load base model + LoRA adapter
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-7B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "haowu1234/signal-dsl-generator-lora")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct")

# Generate DSL
messages = [
    {"role": "system", "content": "You are a Signal DSL configuration generator. Generate valid Signal DSL configurations."},
    {"role": "user", "content": "创建一条路由：当用户提问代码问题时，使用 deepseek-coder 模型"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.1, do_sample=True)

response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)

Example Output

Input: "处理代码相关的问题，使用专业模型"

Output:

SIGNAL domain code_domain {
  description: "Code and programming related queries"
}

ROUTE code_route (description = "Route code questions to specialist") {
  PRIORITY 100
  WHEN domain("code_domain")
  MODEL "deepseek-coder" (reasoning = true)
}

Training Details

Stage 1: Supervised Fine-Tuning (SFT)

• Dataset: 10K+ synthetic DSL examples
• Epochs: 3
• Learning rate: 2e-4

Stage 2: Preference Tuning

• Dataset: Preference pairs (correct vs incorrect DSL)
• Method: Contrastive learning

Stage 3: Direct Preference Optimization (DPO)

• Beta: 0.1
• Final metrics:
  • Accuracy: 96.25%
  • Margins: 1.57
  • Loss: 0.27

Limitations

• Generates Signal DSL syntax specifically; not a general-purpose code generator
• Best results with clear, specific natural language descriptions
• May produce verbose configurations for simple requests

Citation

@misc{signal-dsl-generator,
  author = {Signal Router Team},
  title = {Signal DSL Generator: LoRA Fine-tuned Model for DSL Configuration Generation},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/haowu1234/signal-dsl-generator-lora}
}

License

Apache 2.0 - See LICENSE for details.