llm.create_chat_completion(
messages = [
{
"role": "user",
"content": "What is the capital of France?"
}
]
)MIRA-mini v1.0 — Materials Intelligence and Reasoning Agent
MIRA is a domain-specialized chat model fine-tuned from Qwen2.5-1.5B-Instruct to explain material stacks, physical mechanisms, and correlations between material properties and phenomena in condensed matter physics, device engineering, and materials science.
Developed by DeviceAlchemy.ai.
This repo hosts the GGUF (Q4_K_M quantized) version for local inference with Ollama or llama.cpp.
Training code: github.com/DeviceAlchemy/devicealchemy-mira-mini
Model details
- Base model: Qwen/Qwen2.5-1.5B-Instruct (Alibaba, Apache 2.0)
- Fine-tuning method: QLoRA — continued pre-training on raw abstracts, followed by instruction tuning on synthetically generated Q&A pairs derived from those abstracts
- Quantization: Q4_K_M (~940 MB)
- Training data: Openly licensed (CC BY) scientific text dataset containing 120 million words.
- Context length: 4096 tokens (inference); trained on 384-token blocks
Intended use
MIRA-mini is designed to:
- Explain why specific material stacks are predicted to exhibit a given phenomenon
- Describe physical mechanisms at material interfaces (spin-orbit coupling, exchange interactions, band hybridization, etc.)
- Answer domain questions in condensed matter physics, materials science, and device & circuit engineering
It is not a substitute for primary literature, peer review, or experimental validation. Outputs may contain errors, omissions, or hallucinations and should be verified against original research.
How to use (Ollama)
ollama pull hf.co/DeviceAlchemy/devicealchemy-mira-mini
ollama run hf.co/DeviceAlchemy/devicealchemy-mira-mini
Or download the GGUF directly and build a local Modelfile:
FROM ./mira-mini-q4_k_m.gguf
TEMPLATE """<|im_start|>system
{{ .System }}<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""
SYSTEM """You are MIRA, Materials Intelligence and Reasoning Agent. You have been trained on a large corpus of scientific abstracts on condensed matter physics, device engineering, and materials science. You explain material stack predictions, physical mechanisms, and correlations between material properties and phenomena. Be precise, scientific, and concise."""
PARAMETER temperature 0.8
PARAMETER top_p 0.9
PARAMETER top_k 40
PARAMETER repeat_penalty 1.5
PARAMETER num_ctx 4096
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|im_start|>"
ollama create mira-mini -f Modelfile
ollama run mira-mini
Training pipeline
Full training code (4-step pipeline: data prep → continued pre-training → chat fine-tuning → GGUF export) is available at: github.com/DeviceAlchemy/devicealchemy-mira-mini
Limitations
- This is a 1.5B parameter model — it has narrower general-purpose reasoning ability than larger models, and is specifically tuned toward the materials science and device engineering domain.
- Training data consists of abstracts only (not full papers), so depth on any single topic is bounded by what's typically conveyed in an abstract.
- Q4_K_M quantization trades a small amount of precision for a ~3x smaller file size and faster local inference; the full-precision merged model is referenced in the GitHub repo for those who want F16 or further quantization options.
Disclaimer
MIRA-mini outputs are for informational and exploratory purposes only and should not be considered professional, scientific, medical, legal, or financial advice. Always verify critical information against original research papers and conventional theoretical/experimental means. DeviceAlchemy.ai is not affiliated with, endorsed by, or partnered with any academic publishers.
Author
Shehrin Sayed, Ph.D. — Founder, DeviceAlchemy.ai
- Downloads last month
- 39
4-bit
# !pip install llama-cpp-python from llama_cpp import Llama llm = Llama.from_pretrained( repo_id="DeviceAlchemy/devicealchemy-mira-mini", filename="mira-mini-q4_k_m.gguf", )