Qwen2.5-Math-7B-Instruct-4bit

Qwen2.5-Math-7B-Instruct-4bit

Model Description

Qwen2.5-Math-7B-Instruct-4bit is a 4-bit quantized version of the Qwen/Qwen2.5-Math-7B-Instruct model using GPTQ quantization (W4A16 - 4-bit weights, 16-bit activations).

This model is optimized to:

• Reduce model size by ~75% compared to the original model
• Reduce GPU memory requirements during inference
• Increase inference speed
• Maintain high accuracy for mathematical tasks

Model Details

• Developed by: Community
• Model type: Causal Language Model (Quantized)
• Language(s): English, Mathematics
• License: MIT
• Finetuned from model: Qwen/Qwen2.5-Math-7B-Instruct
• Quantization method: GPTQ (W4A16) via LLM Compressor
• Calibration dataset: GSM8K (256 samples)

Model Sources

• Base Model: Qwen/Qwen2.5-Math-7B-Instruct
• Quantization Tool: vLLM LLM Compressor

Uses

Direct Use

This model is designed for direct use in mathematical and reasoning tasks, including:

• Solving arithmetic, algebra, and geometry problems
• Mathematical reasoning and proofs
• Analyzing and explaining mathematical concepts
• Educational mathematics support

Example Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "your-username/qwen2.5-math-7b-instruct-4bit"
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    device_map="auto",
    dtype="float16",
    trust_remote_code=True,
    low_cpu_mem_usage=False,  # Important for compressed models
)

# Create prompt
prompt = "<|im_start|>user\nSolve for x: 3x + 5 = 14<|im_end|>\n<|im_start|>assistant\n"

# Generate
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200, do_sample=False)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Downstream Use

This model can be further fine-tuned for specific mathematical tasks or integrated into educational applications.

Out-of-Scope Use

This model is NOT designed for:

• Generating harmful or inappropriate content
• Use in applications requiring absolute accuracy (such as critical financial calculations)
• Tasks unrelated to mathematics or reasoning

Bias, Risks, and Limitations

Limitations

• The model has been quantized and may have slightly lower accuracy compared to the original model
• May encounter errors with some complex problems or edge cases
• Model was primarily trained on English data

Recommendations

Users should:

• Verify results for important mathematical problems
• Use the original model (full precision) if maximum accuracy is required
• Understand that quantization may affect some tasks

How to Get Started with the Model

Installation

pip install transformers torch accelerate

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "your-username/qwen2.5-math-7b-instruct-4bit"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    dtype="float16",
    trust_remote_code=True,
    low_cpu_mem_usage=False,
)

# Use the model
prompt = "<|im_start|>user\nWhat is 2+2?<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Quantization Procedure

The model was quantized using:

• Method: GPTQ (W4A16)
• Tool: vLLM LLM Compressor
• Calibration dataset: GSM8K (256 samples)
• Max sequence length: 2048 tokens
• Target layers: All Linear layers except lm_head

Quantization Hyperparameters

• Scheme: W4A16 (4-bit weights, 16-bit activations)
• Block size: 128
• Dampening fraction: 0.01
• Calibration samples: 256

Evaluation

Testing Data

The model was evaluated on the GSM8K test set.

Metrics

• Accuracy: Measured on GSM8K test set
• Model size: ~3.5GB (compared to ~14GB of the original model)
• Compression ratio: ~75% reduction
• Memory usage: Significantly reduced compared to the original model

Results

The compressed model maintains high accuracy for mathematical tasks while significantly reducing size and memory requirements.

Technical Specifications

Model Architecture

• Base Architecture: Qwen2.5 (Transformer-based)
• Parameters: 7B (quantized to 4-bit)
• Context Length: 8192 tokens (original model), 2048 tokens (optimized for quantization)
• Quantization: GPTQ W4A16

Compute Infrastructure

Hardware

• Training/Quantization: NVIDIA RTX 3060 12GB (or equivalent)
• Minimum Inference: GPU with at least 8GB VRAM

Software

• Quantization Tool: vLLM LLM Compressor
• Framework: PyTorch, Transformers
• Python: >=3.12

Citation

If you use this model, please cite:

Base Model:

@article{qwen2.5,
  title={Qwen2.5: A Large Language Model for Mathematics},
  author={Qwen Team},
  year={2024}
}

Quantization Method:

@article{gptq,
  title={GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers},
  author={Frantar, Elias and Ashkboos, Saleh and Hoefler, Torsten and Alistarh, Dan},
  journal={arXiv preprint arXiv:2210.17323},
  year={2022}
}

Model Card Contact

To report issues or ask questions, please open an issue on the repository.

Acknowledgments

• Qwen Team for the original Qwen2.5-Math-7B-Instruct model
• vLLM team for the LLM Compressor tool
• Hugging Face for infrastructure and support