Christine-HiAiPerf/saul-7b-canadian-tax-dpo

Saul-7B Canadian Tax Law (DPO-Aligned)

This is a DPO-aligned version of Equall/Saul-7B-Instruct-v1 fine-tuned on Canadian Tax Law with Direct Preference Optimization.

Model Description

• Base Model: Saul-7B-Instruct-v1 (Legal domain specialist)
• Training: SFT (day3) + DPO alignment (day6)
• Domain: Canadian Income Tax Act & Excise Tax Act
• Training Data: 10,180 Q&A pairs + 900 preference pairs
• Hardware: NVIDIA RTX A6000 48GB
• Training Time: ~3-4 hours total (SFT + DPO)

Training Details

SFT Phase (day3)

• Supervised fine-tuning on Canadian tax Q&A pairs
• LoRA adapter (rank=16, alpha=32)
• 3 epochs, FP16 precision

DPO Phase (day6)

• Direct Preference Optimization for alignment
• Trained on preference pairs (chosen vs rejected responses)
• Beta=0.1, 2 epochs
• Optimized for RTX A6000 with TF32

Improvements over SFT-only

DPO alignment provides:

• ✅ Fewer hallucinations (more accurate citations)
• ✅ More precise legal terminology
• ✅ Better instruction following
• ✅ More helpful and relevant responses

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model
model = AutoModelForCausalLM.from_pretrained(
    "Christine-HiAiPerf/saul-7b-canadian-tax-dpo",
    torch_dtype=torch.float16,
    device_map="auto"
)

tokenizer = AutoTokenizer.from_pretrained("Christine-HiAiPerf/saul-7b-canadian-tax-dpo")

# Generate response
question = "What is Section 118 of the Income Tax Act?"
prompt = f"<s>[INST] {question} [/INST]"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=300)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(response.split("[/INST]")[1])

Training Configuration

# DPO Settings
BATCH_SIZE = 4
GRADIENT_ACCUMULATION = 4
LEARNING_RATE = 5e-7
EPOCHS = 2
BETA = 0.1

# Optimization (RTX A6000)
- FP16 precision
- TF32 enabled
- Fused AdamW optimizer
- No gradient checkpointing

Limitations

• Trained specifically on Canadian federal tax law
• May not cover provincial tax differences
• Not a replacement for professional tax advice
• Based on training data available up to training time

Citation

If you use this model, please cite:

@misc{saul-7b-canadian-tax-dpo,
  author = {Christine-HiAiPerf},
  title = {Saul-7B Canadian Tax Law (DPO-Aligned)},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/Christine-HiAiPerf/saul-7b-canadian-tax-dpo}
}

License

Apache 2.0 (same as base Saul-7B model)

Acknowledgments

• Base model: Equall/Saul-7B-Instruct-v1
• Training framework: HuggingFace Transformers + TRL (DPO)
• LoRA: PEFT library