qwen-25-14b-instruct-quantum-physics

This model is a fine-tuned version of Qwen/Qwen2.5-14B-Instruct using LoRA (Low-Rank Adaptation) on a quantum physics dataset.

Evaluation

Metric	Base Model	Fine-Tuned (SFT)	Fine-Tuned (latest)
Overall Accuracy	24.0%	41.4%	53.7%
Factual Accuracy	—	—	55.0
Completeness	—	—	51.0
Technical Precision	—	—	54.3

Evaluated on BoltzmannEntropy/QuantumLLMInstruct with RAG-augmented judging (Semantic Scholar, 5 papers per question).

Available Formats

• GGUF (Q4_K_M): qwen-25-14b-quantum-physics-q4_k_m.gguf — 8.4 GB, quantized for efficient inference
• GGUF (FP16): _temp_merged_qwen-25-14b-instruct-14b-quantum-physics-20260125-007.fp16.gguf — full precision

Usage

Using GGUF (with llama.cpp, Ollama, LM Studio, etc.)

# Download the quantized GGUF
huggingface-cli download Kylan12/qwen-25-14b-instruct-quantum-physics qwen-25-14b-quantum-physics-q4_k_m.gguf

# Use with llama.cpp
./llama.cpp/build/bin/llama-cli -m qwen-25-14b-quantum-physics-q4_k_m.gguf -p "Your prompt here"

Using HuggingFace Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Kylan12/qwen-25-14b-instruct-quantum-physics")
tokenizer = AutoTokenizer.from_pretrained("Kylan12/qwen-25-14b-instruct-quantum-physics")

prompt = "Calculate the expectation value of the Pauli Z operator for a qubit in the state |+⟩"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
print(tokenizer.decode(outputs[0]))

Training Details

• Base Model: Qwen/Qwen2.5-14B-Instruct
• Training Method: LoRA (Low-Rank Adaptation)
• Quantization: 4-bit NF4 via bitsandbytes
• LoRA Rank: 16
• LoRA Alpha: 16
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Limitations

This model inherits the limitations of the base Qwen2.5-14B-Instruct model and may have additional domain-specific limitations due to the fine-tuning dataset.

License

This model is released under the Apache 2.0 license, consistent with the base Qwen model.