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

python

# 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:

bibtex

@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