README.md

metadata

license: ms-pl
tags:
  - merge
model-index:
  - name: Orca2-13B-selfmerge-39B
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: AI2 Reasoning Challenge (25-Shot)
          type: ai2_arc
          config: ARC-Challenge
          split: test
          args:
            num_few_shot: 25
        metrics:
          - type: acc_norm
            value: 60.84
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=vmajor/Orca2-13B-selfmerge-39B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: HellaSwag (10-Shot)
          type: hellaswag
          split: validation
          args:
            num_few_shot: 10
        metrics:
          - type: acc_norm
            value: 79.84
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=vmajor/Orca2-13B-selfmerge-39B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU (5-Shot)
          type: cais/mmlu
          config: all
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 60.32
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=vmajor/Orca2-13B-selfmerge-39B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: TruthfulQA (0-shot)
          type: truthful_qa
          config: multiple_choice
          split: validation
          args:
            num_few_shot: 0
        metrics:
          - type: mc2
            value: 56.38
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=vmajor/Orca2-13B-selfmerge-39B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: Winogrande (5-shot)
          type: winogrande
          config: winogrande_xl
          split: validation
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 76.87
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=vmajor/Orca2-13B-selfmerge-39B
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GSM8k (5-shot)
          type: gsm8k
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 39.2
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=vmajor/Orca2-13B-selfmerge-39B
          name: Open LLM Leaderboard

This model is a result of merging three Orca2-13B models with itself using 'mergekit-legacy'. Merge parameters were passthrough for one and --weight 0.5 --density 0.5 for the second merge.

This merged model showed marginal improvement in perplexity scores:

The perplexity for Orca-2-13b is: 7.595028877258301 The perplexity for orca2-26B-self-merge is: 7.550178050994873 The perplexity for orca2-39B-self-merge is: NC

The following table summarizes the model performance across a range of benchmarks:

Model	Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande	GSM8K
microsoft/Orca-2-13b	58.64	60.67	79.81	60.37	56.41	76.64	17.97
vmajor/Orca2-13B-selfmerge-26B	62.24	60.84	79.84	60.32	56.38	76.87	39.2
vmajor/Orca2-13B-selfmerge-39B	62.24	60.84	79.84	60.32	56.38	76.87	39.2

Interestingly the GSM8K performance more than doubled with the first self merge. Second self merge resulting in the 39B model did not produce any further gains.

---

license: ms-pl

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	62.24
AI2 Reasoning Challenge (25-Shot)	60.84
HellaSwag (10-Shot)	79.84
MMLU (5-Shot)	60.32
TruthfulQA (0-shot)	56.38
Winogrande (5-shot)	76.87
GSM8k (5-shot)	39.20