suayptalha/FastLlama-3.2-LoRA

These are only LoRA adapters of FastLlama-3.2-1B-Instruct. You should also import the base model in order to use them!

You can use ChatML & Alpaca format.

You can chat with the model via this space.

Overview:

FastLlama is a highly optimized version of the Llama-3.2-1B-Instruct model. Designed for superior performance in constrained environments, it combines speed, compactness, and high accuracy. This version has been fine-tuned using the MetaMathQA-50k section of the HuggingFaceTB/smoltalk dataset to enhance its mathematical reasoning and problem-solving abilities.

Features:

Lightweight and Fast: Optimized to deliver Llama-class capabilities with reduced computational overhead. Fine-Tuned for Math Reasoning: Utilizes MetaMathQA-50k for better handling of complex mathematical problems and logical reasoning tasks. Instruction-Tuned: Pre-trained on instruction-following tasks, making it robust in understanding and executing detailed queries. Versatile Use Cases: Suitable for educational tools, tutoring systems, or any application requiring mathematical reasoning.

Performance Highlights:

Smaller Footprint: The model delivers comparable results to larger counterparts while operating efficiently on smaller hardware. Enhanced Accuracy: Demonstrates improved performance on mathematical QA benchmarks. Instruction Adherence: Retains high fidelity in understanding and following user instructions, even for complex queries.

Loading the Model:

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

base_model_id = "meta-llama/Llama-3.2-1B-Instruct"  # Base model ID
adapter_id = "suayptalha/FastLlama-3.2-LoRA"  # Adapter ID

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

model = PeftModel.from_pretrained(base_model, adapter_id)

# Text generation pipeline
pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a friendly assistant named FastLlama."},
    {"role": "user", "content": "Who are you?"},
]

outputs = pipe(
    messages,
    max_new_tokens=256,
)

print(outputs[0]["generated_text"][-1])

Dataset:

Dataset: MetaMathQA-50k

The MetaMathQA-50k subset of HuggingFaceTB/smoltalk was selected for fine-tuning due to its focus on mathematical reasoning, multi-step problem-solving, and logical inference. The dataset includes:

Algebraic problems Geometric reasoning tasks Statistical and probabilistic questions Logical deduction problems

Model Fine-Tuning:

Fine-tuning was conducted using the following configuration:

Learning Rate: 2e-4

Epochs: 1

Optimizer: AdamW

Framework: Unsloth

License:

This model is licensed under the Apache 2.0 License. See the LICENSE file for details.