yi-34b-w4a16g32 / README.md
NicoNico's picture
Update README.md
d6fa456
|
raw
history blame
2.15 kB

license: apache-2.0

GreenBit Yi

This is GreenBitAI's pretrained 2-bit Yi 34B model with extreme compression yet still strong performance.

Please refer to our Github page for the code to run the model and more information.

Model Description

Model Yi-34B Yi-6B
Bit 16 4 2 16 4 2
GroupSize - 32 8 - 32 8
Model Size (GB) 68.79 19.89 15.508 12.12 4.04 3.32
AVG 70.64 69.7 65.78 60.11 59.14 53.69
Detailed Evaluation
MMLU 76.32 75.42 72 63.24 62.09 58.57
CMMLU 83.65 83.07 78.26 75.53 72.85 64.83
ARC-e 84.42 84.13 81.35 77.23 76.52 73.15
ARC-c 61.77 59.56 57 50.34 48.47 42.58
GAOKAO 82.8 81.37 77.9 72.2 72.87 64.67
GSM8K 67.24 63.61 50.1895 32.52 28.05 16.98
HumanEval 25.6 25 18.9024 15.85 15.85 12.19
BBH 54.3 52.3 49.0401 42.8 41.47 38.12
WinoGrande 78.68 78.53 77.98 70.63 71.19 67.96
PIQA 82.86 82.75 81.72 78.56 79.05 76.55
SIQA 74.46 73.44 71.6 64.53 64.53 57
HellaSwag 83.64 83.02 80.7 74.91 73.27 69.07
OBQA 91.6 90.8 84 85.4 82.6 79.4
CSQA 83.37 83.05 81.5725 76.9 75.43 69.53
TriviaQA 81.52 80.73 75.8163 64.85 61.75 50.5
SquAD 92.46 91.12 89.3624 88.95 88.39 84.22
BoolQ 88.25 88.17 80.74 76.23 77.1 75.5
MBPP 41 39.68 31.4815 26.32 25.13 16
QUAC 48.61 47.43 46.5406 40.92 40.16 37.83
Lambda 73.18 73.39 72.83 67.74 67.8 61.8
NaturalQuestion 27.67 27.21 22.4994 16.69 17.42 11.05