Merged LLaMA Model

This is a merged version of the LLaMA2-13b model based on hyperboloid projections. The model retains 31 layers with significant performance retention across all benchmarks.

Downloads last month: 1

Safetensors

Model size

10B params

Tensor type

F16

·

Model tree for namannn/llama2-13b-hyperbolic-cluster-pruned

Base model

meta-llama/Llama-2-13b

Finetuned

(6)

this model

Quantizations

1 model

Space using namannn/llama2-13b-hyperbolic-cluster-pruned 1

Evaluation results

Accuracy on HellaSwag
OpenCompass

66.660
Accuracy on PIQA
OpenCompass

72.140
Perplexity on WSC
OpenCompass

60.380
Accuracy on WSC
OpenCompass

38.990
Accuracy on CSQA
OpenCompass

54.360
Accuracy on MMLU
OpenCompass

54.760
Accuracy on RACE
OpenCompass

53.890
Accuracy on RACE
OpenCompass

55.290