newmindai/Llama-3.1-8B-Instruct-environmental-alpaca

<div style="display: flex; justify-content: center; flex-wrap: wrap; gap: 15px; align-items: flex-start;">

<div style="padding: 10px; border-radius: 8px; background-color: #000; color-scheme: light dark; margin-top: 40px;"> <img src="https://cdn-uploads.huggingface.co/production/uploads/683d4880e639f8d647355997/iA0QsWk-k-fQ7CCt65znl.png" style="width: 100%; max-width: 250px; height: 40px; filter: invert(0) brightness(1) contrast(1);" /> </div>

<img src="https://cdn-uploads.huggingface.co/production/uploads/683d4880e639f8d647355997/mqbOdFfrC7KjDZbQlLFFj.png" style="width: 100%; max-width: 300px; height: auto;" />

<img src="https://cdn-uploads.huggingface.co/production/uploads/683d4880e639f8d647355997/VGnh14pYg-SXSaEt640qz.png" style="width: 100%; max-width: 150px; height: 100px; margin-top: 35px;" />

</div>

Model Card

This document describes a parameter-efficient fine-tuning setup using LoRA on the EuroHPC Karolina system. Axolotl provides flexible orchestration and Unsloth supplies optimized kernels for high-throughput training on the newmindai/EuroHPC-Legal dataset. This model is fine-tuned using Parameter-Efficient Fine-Tuning (PEFT) with LoRA (Low-Rank Adaptation) on the EuroHPC dataset, specifically the environmental subset. The fine-tuning leverages the Axolotl framework for orchestration and Unsloth library for optimized training kernels.

Hyperparameters

• —LoRA Rank: 16
• —LoRA Alpha: 32
• —LoRA Dropout: 0.05
• —Learning Rate: 3×10⁻⁵ with cosine scheduling
• —Training Epochs: 3 per domain
• —Batch Size: Optimized for A100 memory capacity

Architecture

• —Base Model: Llama-3.1-8B-Instruct (Meta)
• —Fine-tuning Method: LoRA (Low-Rank Adaptation)
• —Parameter Efficiency: Only trainable LoRA parameters, frozen base model
• —Model Size: 8B parameters (base) + LoRA adapters

Hardware and Software

• —Orchestration: Axolotl framework
• —Acceleration: Unsloth library
• —Backend: PyTorch with CUDA support
• —System: EuroHPC Karolina supercomputer
• —GPUs: NVIDIA A100 (8 × 40 GB per node, 320 GB HBM2 total)
• —Utilization: 85–90% GPU and memory efficiency
• —Total Compute: ~600 GPU hours

Data

Input Format

The dataset follows the Alpaca format with three key fields:

{
  "instruction": "Task description or question",
  "input": "Additional context or input data",
  "output": "Expected response or answer"
}

Dataset: newmindai/EuroHPC-Legal (environmental subset)

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load base model and tokenizer
base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "newmindai/Llama-3.1-8B-Instruct-environmental-alpaca")

# Format input according to Alpaca format
def format_prompt(instruction, input_text=""):
    if input_text:
        return f"### Instruction:\n{instruction}\n\n### Input:\n{input_text}\n\n### Response:\n"
    else:
        return f"### Instruction:\n{instruction}\n\n### Response:\n"

# Example usage
prompt = format_prompt("Explain the benefits of regular exercise")
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Acknowledgments

This research was supported by the EuroHPC Joint Undertaking (EuroHPC JU) under the Benchmark Access grant agreement No EHPC-BEN-2024B11-003. The authors gratefully acknowledge the computational resources provided by the IT4Innovations National Supercomputing Center (Czech Republic) on the Karolina supercomputer, made available through the EuroHPC JU.

Citation

@article{newmind2025,
  title={Tailoring AI for Turkish Law: Domain-Specific Fine-Tuning of Small Language Models for Legal Expertise},
  author={New Mind AI Team},
  journal={Procedia Computer Science},
  year={2025},
  volume={239},
  doi={10.1016/j.procs.2025.08.239},
  note={Available online 23 September 2025, Version of Record 23 September 2025}
}