README.md

metadata

license: apache-2.0
language:
  - en
pipeline_tag: text-generation
tags:
  - mistral
  - fine-tuned
  - lora
  - unsloth
base_model: unsloth/Mistral-Nemo-Instruct-2407-bnb-4bit

aLLoyM: A large language model for alloy phase diagram prediction

NEW!! Check out our App on Huggingface Spaces. You can chat with aLLoyM or interactively draw generated phase diagrams directly in your browser—no local GPU required.

Our preprint on arXiv

Data used to train the model are here.

Model Details

• Model Name: aLLoyM
• Base Model: unsloth/Mistral-Nemo-Instruct-2407-bnb-4bit
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Training Framework: Unsloth

Usage

Basic Usage

from unsloth import FastLanguageModel
import torch
from huggingface_hub import login

# Authenticate with Hugging Face
login('YOUR_HF_TOKEN')

# Model configuration
max_seq_length = 2048
dtype = torch.bfloat16
load_in_4bit = True

print("Loading model from Hugging Face...")
# Load model and tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name='Playingyoyo/aLLoyM',
    max_seq_length=max_seq_length,
    dtype=dtype,
    load_in_4bit=load_in_4bit,
)
FastLanguageModel.for_inference(model)
print("Model loaded successfully!")

# Define the question
question = "What phases form when Arsenic (40%) + Platinum (60%) are mixed at 400 K?" # Replace here with your own question

# Create prompt template
prompt = f"""### Instruction:
You are an expert in phase diagrams, thermodynamics, and materials science, specializing in binary alloy systems.

### Input:
{question}

### Output:
"""

# Tokenize input
inputs = tokenizer(
    [prompt],
    return_tensors='pt',
    truncation=True
).to('cuda')

# Generate response
print(f"\nGenerating response for: '{question}'")
with torch.no_grad():
    outputs = model.generate(
        **inputs, 
        max_new_tokens=512,
        use_cache=True,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id
    )

# Decode and extract the generated response
full_output = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]

# Extract only the generated part after "### Output:"
if "### Output:" in full_output:
    generated_response = full_output.split("### Output:")[1].strip()
else:
    generated_response = full_output.strip()

print(f"\nAnswer:")
print("=" * 50)
print(generated_response)
print("=" * 50)

Question Samples

aLLoyM was trained using a standardized prompt template for consistency, which may make it sensitive to variations in prompt formulation. Users should be aware that rephrasing questions or changing the input format may affect prediction qual- ity. We encourage the community to experiment with different prompting approaches and share effective strategies.

Training Configuration

• Learning Rate: 2e-4
• Batch Size: 16 (per device)
• Gradient Accumulation Steps: 4
• LoRA Rank: 16
• LoRA Alpha: 16
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

License

Apache 2.0

Citation

If you use this model, please cite:

@misc{aLLoyM,
  title=,
  author=,
  year=,
  publisher=,
  howpublished=
}