karunakant/vedalytics-gemma4E4B-chemistry

🌟 Vedalytics Gemma 4 (E4B) — Chemistry & Spectroscopy Specialized Model

An expert language model fine-tuned for chemistry, NMR, IR, mass spectrometry, and spectroscopy, optimized for enterprise efficiency on both CPU and GPU.

---

🎯 Model Overview & Capabilities

Vedalytics Gemma 4 is a 4-billion parameter transformer model built on Google's Gemma architecture (Gemma 4-E4B), enhanced with high-precision quantization and enterprise-grade fine-tuning.

The model is specialized for:

• Organic Chemistry, Inorganic Chemistry, Analytical Chemistry
• Nuclear Magnetic Resonance (¹H, ¹³C, 2D NMR)
• Infrared (IR) and Raman Spectroscopy
• UV-Visible (UV-Vis) and Fluorescence Spectroscopy
• Mass Spectrometry (MS, LC-MS, GC-MS)
• Reaction Mechanisms, Synthesis, and Spectroscopy Interpretation

Key features:

• Domain-Specific Expertise: Trained on 5,000+ chemistry-focused Q&A pairs with industry terminology.
• Guardrails & Safety: Automatically rejects out-of-domain queries (e.g., general knowledge, history, programming) to maintain scientific integrity.
• Optimized for Inference:
  • CPU: 4-bit or 8-bit quantization for high performance on standard Intel/AMD hardware.
  • GPU: Full 16-bit or 32-bit precision with Tensor Core acceleration on NVIDIA/DGX systems.
  • Mac Metal: Metal Performance Shaders support for optimal Apple Silicon (M1/M2/M3/M4) performance.

The fine-tuned model is constrained to only respond to questions regarding Chemistry and Spectroscopy (organic/inorganic/analytical chemistry, NMR, IR, UV-Vis, Mass Spectrometry, Raman, etc.) and politely reject out-of-domain queries.

---

🚀 How to Get Started

Using Merged Model (PyTorch / Transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "./merged_model"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

prompt = "Explain chemical shift in 1H NMR spectroscopy."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=256)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Using GGUF with Ollama or llama.cpp

# Run with Ollama
ollama run gemma-chem

# Run with llama.cpp
./llama-cli -m ./output_gguf/vedalytics-gemma4E4B-chemistry.gguf -t 8 -p "What is the IR peak for an alkyne group?"

---

🌐 Cross-Platform Deployment Guide (CPU, Mac Mini & DGX Spark)

1. Standard CPU Deployment (x86_64 / Intel / AMD)

• Compatibility: 100% Supported (No GPU required).
• Execution: The exported GGUF model uses llama.cpp AVX2 / AVX-512 CPU thread vectorization.
• Commands:
  • Ollama CPU: ollama run gemma-chem
  • llama.cpp CLI:

    ./llama-cli -m ./output_gguf/vedalytics-gemma4E4B-chemistry.gguf -t 8 -p "Explain chemical shift in 1H NMR."
    

2. Apple Silicon Mac Mini (M1 / M2 / M3 / M4)

• Compatibility: 100% Supported with Metal & Unified Memory Acceleration.
• Performance: High inference speed (~40–70 tokens/sec) utilizing macOS Metal Performance Shaders (MPS).
• Commands:
  • Ollama Mac:

    ollama create gemma-chem -f Modelfile
    ollama run gemma-chem
    

  • llama.cpp (Metal):

    ./llama-cli -m ./output_gguf/vedalytics-gemma4E4B-chemistry.gguf -ngl 99 -p "What is the IR peak for an alkyne?"
    

3. NVIDIA DGX / Spark GPU Clusters

• Compatibility: 100% Supported with Enterprise CUDA / Tensor Core acceleration.
• Deployment:
  • vLLM / Hugging Face: Deploy merged_model/ for multi-user REST API endpoints:

    from transformers import AutoModelForCausalLM, AutoTokenizer
    model = AutoModelForCausalLM.from_pretrained("./merged_model", device_map="cuda")
    

  • Ollama / llama.cpp (CUDA): Load .gguf file with 100% CUDA layer offloading (-ngl 99).

---

📊 Cross-Platform Feature Comparison Matrix

Environment	Acceleration Engine	Est. Speed	Recommended Format
Standard CPU	CPU AVX2 / AVX-512	~15 – 30 tokens/sec	GGUF (f16 or q4_k_m)
Mac Mini (Apple Silicon)	Apple Metal (Unified RAM)	~40 – 70 tokens/sec	GGUF (f16 or q8_0)
NVIDIA DGX / Spark Cluster	CUDA / Tensor Cores	~100+ tokens/sec	Merged Model (merged_model) or GGUF

---

🛠️ Framework Versions

• PEFT: 0.20.0
• TRL: 1.10.0
• Transformers: 5.16.1
• PyTorch: 2.11.0