README.md

metadata

license: mit
base_model:
  - meta-llama/Llama-2-13b
pipeline_tag: text-generation
tags:
  - chemistry
  - biology
  - finance
  - legal
  - music
  - art
  - code
  - climate
  - medical
  - quantized
library_name: transformers
metrics:
  - perplexity
model-index:
  - name: LLaMa2-13b-merged-clusters
    results:
      - task:
          type: commonsense-reasoning
        dataset:
          name: HellaSwag
          type: HellaSwag
        metrics:
          - name: Accuracy
            type: accuracy
            value: 66.66
        source:
          name: OpenCompass
          url: https://opencompass.org
      - task:
          type: commonsense-reasoning
        dataset:
          name: PIQA
          type: PIQA
        metrics:
          - name: Accuracy
            type: accuracy
            value: 72.14
        source:
          name: OpenCompass
          url: https://opencompass.org
      - task:
          type: coreference-resolution
        dataset:
          name: WSC
          type: WSC
        metrics:
          - name: Perplexity
            type: perplexity
            value: 60.38
        source:
          name: OpenCompass
          url: https://opencompass.org
      - task:
          type: coreference-resolution
        dataset:
          name: WSC
          type: WSC
        metrics:
          - name: Accuracy
            type: accuracy
            value: 38.99
        source:
          name: OpenCompass
          url: https://opencompass.org
      - task:
          type: multiple-choice-question-answering
        dataset:
          name: CSQA
          type: CSQA
        metrics:
          - name: Accuracy
            type: accuracy
            value: 54.36
        source:
          name: OpenCompass
          url: https://opencompass.org
      - task:
          type: multi-task-evaluation
        dataset:
          name: MMLU
          type: MMLU
        metrics:
          - name: Accuracy
            type: accuracy
            value: 54.76
        source:
          name: OpenCompass
          url: https://opencompass.org
      - task:
          type: multiple-choice-question-answering
        dataset:
          name: RACE
          type: RACE-high
        metrics:
          - name: Accuracy
            type: accuracy
            value: 53.89
        source:
          name: OpenCompass
          url: https://opencompass.org
      - task:
          type: multiple-choice-question-answering
        dataset:
          name: RACE
          type: RACE-middle
        metrics:
          - name: Accuracy
            type: accuracy
            value: 55.29
        source:
          name: OpenCompass
          url: https://opencompass.org

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.