Llama-3.1-8B-Orca-Structured-LoRA

This is a custom-trained LoRA adapter for Meta-Llama-3.1-8B-Instruct. It has been fine-tuned to excel at step-by-step mathematical reasoning and to respond with a polite, highly structured, and logically organized tone.

Model Highlights

• Base Model: Llama-3.1-8B-Instruct
• Primary Skills: Complex math word problems, logical reasoning, structured explanations
• Output Style: Helpful, and well-organized
• Training Hardware: Single Tesla T4 GPU with Unsloth's 4-bit quantization optimization

---

Training Data

The model combines two carefully curated datasets. A balanced subset of 16,000 examples was used:

1. mlabonne/FineTome-100k: 8,000 examples - Teaches polite conversational tone and deep reasoning chain formatting.
2. microsoft/orca-math-word-problems-200k: 8,000 examples - Provides step-by-step mathematical problem-solving capabilities.

---

How to Use (Inference)

Run this model easily with the unsloth library. It automatically downloads the base model and applies this LoRA adapter.

1. Install Dependencies

pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
pip install --no-deps xformers trl peft accelerate bitsandbytes

2. Run the Model

from unsloth import FastLanguageModel

# Load model and tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "x0root/Llama-3.1-8B-Orca-Structured-LoRA",
    max_seq_length = 1024,
    dtype = None,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

# Create prompt
messages = [
    {"role": "system", "content": "You are a highly intelligent, polite AI assistant. Always think step-by-step and structure your answers beautifully."},
    {"role": "user", "content": "A store sells apples for $1.20 each and bananas for $0.50 each. If I buy 4 apples and 6 bananas, and I pay with a $20 bill, how much change should I receive? Please explain your reasoning clearly."}
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize = True,
    add_generation_prompt = True,
    return_tensors = "pt",
).to("cuda")

# Generate response
outputs = model.generate(
    input_ids = inputs,
    max_new_tokens = 512,
    use_cache = True,
    temperature = 0.3,
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response.split("assistant\n")[-1])

---

Technical Specifications

Training Hyperparameters

Fine-tuned using Supervised Fine-Tuning (SFT) with the following configuration:

• LoRA Rank (r): 16
• LoRA Alpha: 16
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Max Sequence Length: 1024 tokens
• Batch Size: 1 (per device)
• Gradient Accumulation Steps: 8
• Effective Batch Size: 8
• Learning Rate: 2e-4
• Optimizer: paged_adamw_8bit
• Max Steps: 2000
• Warmup Steps: 50
• Weight Decay: 0.01

Frameworks

• Unsloth (2x faster, memory-efficient training)
• Hugging Face Transformers & TRL
• PyTorch