RichardErkhov/IntelLabs_-_sqft-sparsepeft-mistral-7b-v0.3-50-math-heu-gguf

Quantization made by Richard Erkhov.

Github

Discord

Request more models

sqft-sparsepeft-mistral-7b-v0.3-50-math-heu - GGUF

• Model creator: https://huggingface.co/IntelLabs/
• Original model: https://huggingface.co/IntelLabs/sqft-sparsepeft-mistral-7b-v0.3-50-math-heu/

Name	Quant method	Size
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q2_K.gguf	Q2_K	2.54GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.IQ3_XS.gguf	IQ3_XS	2.82GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.IQ3_S.gguf	IQ3_S	2.97GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q3_K_S.gguf	Q3_K_S	2.95GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.IQ3_M.gguf	IQ3_M	3.06GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q3_K.gguf	Q3_K	3.28GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q3_K_M.gguf	Q3_K_M	3.28GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q3_K_L.gguf	Q3_K_L	3.56GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.IQ4_XS.gguf	IQ4_XS	3.68GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q4_0.gguf	Q4_0	3.83GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.IQ4_NL.gguf	IQ4_NL	3.87GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q4_K_S.gguf	Q4_K_S	3.86GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q4_K.gguf	Q4_K	4.07GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q4_K_M.gguf	Q4_K_M	4.07GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q4_1.gguf	Q4_1	4.24GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q5_0.gguf	Q5_0	4.66GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q5_K_S.gguf	Q5_K_S	4.66GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q5_K.gguf	Q5_K	4.78GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q5_K_M.gguf	Q5_K_M	4.78GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q5_1.gguf	Q5_1	5.07GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q6_K.gguf	Q6_K	5.54GB
sqft-sparsepeft-mistral-7b-v0.3-50-math-heu.Q8_0.gguf	Q8_0	7.17GB

Original model description:

language: en license: apache-2.0 library_name: transformers

SQFT Fine-tuned Model: sqft-sparsepeft-mistral-7b-v0.3-50-math-heu

• Base Model: IntelLabs/sqft-mistral-7b-v0.3-50-base
• Sparsity: 50%
• Quantization: No
• Finetune Method: SQFT + SparsePEFT
• 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

MODEL_NAME=IntelLabs/sqft-sparsepeft-mistral-7b-v0.3-50-math-heu
OUTPUT_DIR=./results
python eval/evaluate_math.py --base_model_path ${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
• 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