kalkiai3000/yi34b-quantum-finetuned-subject-matter

Yi-34B Quantum Fine-tuned with Subject Matter Knowledge

Model Overview

This model is a fine-tuned version of Yi-1.5-34B-Chat specifically optimized for quantum physics instruction-following tasks. The model has been enhanced with dense domain and subdomain knowledge through structured fine-tuning on quantum physics problems with explicit subject matter context.

Key Differentiators

🧠 Subject Matter Contextualization

Unlike standard quantum AI models, this model incorporates explicit subject matter knowledge in its training process. Each training example includes:

• Domain: Primary field (e.g., "Quantum Physics", "Quantum Computing")
• Subdomain: Specific area (e.g., "Quantum Mechanics", "Quantum Algorithms", "Quantum Information Theory")
• Subject Knowledge: Contextual background information that guides reasoning
• Problem: The specific question or task
• Solution: Expert-level response

This structured approach enables the model to:

• Maintain contextual awareness across different quantum physics subfields
• Provide more accurate and domain-appropriate responses
• Demonstrate deeper understanding of subject matter relationships

🎯 Specialized Training Methodology

The model was fine-tuned using QLoRA (Quantized Low-Rank Adaptation) with the following specifications:

• Base Model: Yi-1.5-34B-Chat (34 billion parameters)
• Fine-tuning Method: QLoRA with 4-bit quantization
• LoRA Configuration:
  • Rank (r): 16
  • Alpha: 32
  • Dropout: 0.1
  • Target modules: q_proj, v_proj, k_proj, o_proj, gate_proj, up_proj, down_proj
• Training Hardware: NVIDIA A100-80GB
• Memory Efficiency: ~35-40GB VRAM usage during training

Dataset

Source Dataset: kalkiai3000/quantum-llm-instruct-subject-knowledge

This dataset contains structured quantum physics problems with the following schema:

{
  "problem": "Quantum physics question or task",
  "main_domain": "Primary field classification", 
  "sub_domain": "Specific subdomain",
  "subject_knowledge": "Relevant background knowledge",
  "solution": "Expert-level solution",
  "uuid": "Unique identifier",
  "timestamp": "Creation timestamp",
  "model_name": "Source model information",
  "solution_model_name": "Solution generation model"
}

Training Format

Each training example follows this instruction template:

Problem: [Quantum physics question]

Domain: [main_domain]  
Subdomain: [sub_domain]
Subject Knowledge: [relevant context]

Please provide the solution:

This format ensures the model learns to:

1. Consider domain context when reasoning
2. Apply relevant subject matter knowledge
3. Maintain consistency within subdomains
4. Generate contextually appropriate solutions

Training Details

Infrastructure

• GPU: NVIDIA A100-SXM4-80GB
• Framework: PyTorch + Transformers + PEFT + TRL
• Quantization: BitsAndBytesConfig (4-bit NF4)
• Mixed Precision: BFloat16

Hyperparameters

• Learning Rate: 2e-4
• Training Epochs: 3
• Max Steps: 1000
• Batch Size: 1 per device
• Gradient Accumulation: 8 steps
• Warmup Steps: 100
• Evaluation Strategy: Every 100 steps
• Save Strategy: Every 100 steps

Memory Optimization

• Gradient Checkpointing: Enabled
• QLoRA: 4-bit quantization with double quantization
• Device Mapping: Automatic GPU distribution
• Context Length: 2048 tokens

Performance Characteristics

Specialized Capabilities

1. Domain-Aware Reasoning: Incorporates explicit domain knowledge in problem-solving
2. Subdomain Expertise: Maintains accuracy across quantum physics subfields
3. Contextual Problem Solving: Uses subject matter knowledge to guide solutions
4. Structured Output: Provides well-organized, educational responses

Target Applications

• Quantum physics education and tutoring
• Research assistance in quantum domains
• Technical documentation generation
• Quantum algorithm explanation
• Physics problem solving with context

Planned Benchmarking

This model will be systematically evaluated against leading 32B parameter models:

Primary Comparison: Qwen2.5-32B

• Base vs Fine-tuned: Direct comparison with Qwen2.5-32B-Instruct
• Domain Specialization: Quantum physics problem accuracy
• Context Utilization: Effectiveness of subject matter integration

Secondary Comparisons

1. Yi-34B-Chat (base model) - Measuring fine-tuning improvements
2. DeepSeek-V2-Chat (comparable 32B+ model) - Performance benchmarking
3. Mixtral-8x7B-Instruct (MoE architecture comparison)
4. InternLM2.5-20B (efficiency vs capability trade-off)

Evaluation Metrics

• Accuracy: Correctness of quantum physics solutions
• Domain Consistency: Appropriate use of domain/subdomain context
• Knowledge Integration: Effective utilization of subject matter
• Response Quality: Clarity, completeness, and educational value
• Efficiency: Inference speed and memory usage

Benchmark Datasets

• Quantum physics problem sets from academic sources
• Domain-specific evaluation tasks
• Cross-domain generalization tests
• Subject knowledge application assessments

Usage

Loading the Model

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("kalkiai3000/yi34b-quantum-finetuned-subject-matter")
model = AutoModelForCausalLM.from_pretrained(
    "kalkiai3000/yi34b-quantum-finetuned-subject-matter",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

Inference Example

# Format your prompt with domain context
prompt = '''Problem: Explain quantum entanglement and its role in quantum computing.

Domain: Quantum Computing
Subdomain: Quantum Information Theory  
Subject Knowledge: Quantum entanglement creates correlated quantum states between particles regardless of distance.

Please provide the solution:'''

# Generate response
formatted_prompt = f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(formatted_prompt, return_tensors="pt")

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        do_sample=True,
        top_p=0.9
    )

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

Optimal Input Format

For best results, structure your prompts with explicit domain context:

• Problem: Clear question or task
• Domain: Primary field of inquiry
• Subdomain: Specific area within the domain
• Subject Knowledge: Relevant background information

This format leverages the model's specialized training and produces more accurate, contextually appropriate responses.

Technical Specifications

Model Architecture

• Base: Yi-1.5-34B-Chat transformer architecture
• Parameters: 34 billion (base) + LoRA adapters
• Context Window: Up to 4K tokens (training), extensible to base model limits
• Precision: BFloat16 for inference

Fine-tuning Artifacts

• LoRA Adapters: Low-rank adaptation layers
• Quantization: 4-bit NF4 quantization support
• Memory Requirements: ~35-40GB VRAM for inference (with quantization)

Limitations

1. Domain Specificity: Optimized for quantum physics; may not generalize to other scientific domains
2. Context Dependency: Performs best when provided with domain/subdomain context
3. Training Data: Limited to the scope of the source dataset
4. Inference Requirements: Requires significant GPU memory for optimal performance

Citation

If you use this model in your research, please cite:

@misc{yi34b-quantum-finetuned-2025,
  title={Yi-34B Quantum Fine-tuned with Subject Matter Knowledge},
  author={KalkiAI},
  year={2025},
  publisher={Hugging Face},
  journal={Hugging Face Model Hub},
  howpublished={\url{https://huggingface.co/kalkiai3000/yi34b-quantum-finetuned-subject-matter}}
}

License

This model inherits the Apache-2.0 license from the base Yi-1.5-34B-Chat model. The fine-tuning code and datasets may have additional licensing terms.

Acknowledgments

• Base Model: Yi-1.5-34B-Chat by 01-ai
• Training Dataset: quantum-llm-instruct-subject-knowledge by KalkiAI
• Framework: Hugging Face Transformers, PEFT, TRL, BitsAndBytesConfig
• Infrastructure: NVIDIA A100 GPU computing resources

---

Model trained and released by KalkiAI Research Team - Advancing AI for Scientific Discovery