README.md

metadata

language: en
license: apache-2.0
library_name: transformers

SQFT Fine-tuned Model: sqft-qa-sparsepeft-phi-3-mini-4k-50-gptq-cs-heu

• Base Model: IntelLabs/sqft-phi-3-mini-4k-50-base-gptq
• Sparsity: 50%
• Quantization: INT4 (GPTQ)
• Finetune Method: SQFT + QA-SparsePEFT
• Finetune data: winogrande, boolq, openbookqa, hellaswag, piqa, ai2_arc training dataset (83k)
• Sub-Adapter: Heuristic

Evaluation

MODEL_NAME=IntelLabs/sqft-qa-sparsepeft-phi-3-mini-4k-50-gptq-cs-heu
lm_eval --model hf --model_args pretrained=${MODEL_NAME},add_bos_token=True,trust_remote_code=True --tasks piqa,arc_easy,arc_challenge,hellaswag,openbookqa,boolq,winogrande --batch_size auto:4

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
• Low-Rank Adapters Meet Neural Architecture Search for LLM Compression

Citation

@inproceedings{munoz-etal-2024-sqft,
    title = "{SQFT}: Low-cost Model Adaptation in Low-precision Sparse Foundation Models",
    author = "Munoz, Juan Pablo  and
      Yuan, Jinjie  and
      Jain, Nilesh",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-emnlp.749",
    pages = "12817--12832",
}

License

Apache-2.0