SQFT Fine-tuned Model: sqft-phi-3-mini-4k-50-gptq-math-heu-adapter

• Base Model: IntelLabs/sqft-phi-3-mini-4k-50-base-gptq
• Sparsity: 50%
• Quantization: INT4 (GPTQ)
• Finetune Method: SQFT
• Finetune data: 10K instruction-following math reasoning training dataset from LLM-Adapters (math_10k.json)
• Sub-Adapter: Heuristic

Evaluation

git clone https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning.git haaml && cd haaml/SQFT

BASE_MODEL_NAME=IntelLabs/sqft-phi-3-mini-4k-50-base-gptq
ADAPTER_MODEL_NAME=IntelLabs/sqft-phi-3-mini-4k-50-gptq-math-heu-adapter
OUTPUT_DIR=./results
python eval/evaluate_math.py --base_model_path ${BASE_MODEL_NAME} --adapter_model_path ${ADAPTER_MODEL_NAME} --output_dir ${OUTPUT_DIR}

Refer to our repo for the environment information to run this command.

Model Sources

• Repository: https://github.com/IntelLabs/Hardware-Aware-Automated-Machine-Learning/tree/main/SQFT
• Paper: SQFT: Low-cost Model Adaptation in Low-precision Sparse Foundation Models

Citation

@article{munoz2024sqft,
  title = {SQFT: Low-cost Model Adaptation in Low-precision Sparse Foundation Models},
  author={J. Pablo Munoz and Jinjie Yuan and Nilesh Jain},
  journal={The 2024 Conference on Empirical Methods in Natural Language Processing (Findings)},
  year={2024}
}

License

Apache-2.0