README.md

metadata

license: apache-2.0
tags:
  - mergekit
  - merge
  - moe
base_model:
  - mistralai/Mistral-7B-Instruct-v0.2
  - mistralai/Mistral-7B-Instruct-v0.1
model-index:
  - name: mistral-instruct-moe-experimental
    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: 61.01
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=osanseviero/mistral-instruct-moe-experimental
          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: 81.55
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=osanseviero/mistral-instruct-moe-experimental
          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: 58.22
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=osanseviero/mistral-instruct-moe-experimental
          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: 60.4
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=osanseviero/mistral-instruct-moe-experimental
          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: 76.09
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=osanseviero/mistral-instruct-moe-experimental
          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: 31.08
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=osanseviero/mistral-instruct-moe-experimental
          name: Open LLM Leaderboard

Mistral Instruct MoE experimental

This is a merge of pre-trained language models created using mergekit using the mixtral branch.

This is an experimental model and has nothing to do with Mixtral. Mixtral is not a merge of models per se, but a transformer with MoE layers learned during training

This uses a random gate, so I expect not great results. We'll see!

Merge Details

Merge Method

This model was merged using the MoE merge method.

Models Merged

The following models were included in the merge:

• mistralai/Mistral-7B-Instruct-v0.2
• mistralai/Mistral-7B-Instruct-v0.1

Configuration

The following YAML configuration was used to produce this model:

base_model: mistralai/Mistral-7B-Instruct-v0.2
gate_mode: random
dtype: bfloat16
experts:
  - source_model: mistralai/Mistral-7B-Instruct-v0.2
    positive_prompts: [""]
  - source_model: mistralai/Mistral-7B-Instruct-v0.1
    positive_prompts: [""]

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	61.39
AI2 Reasoning Challenge (25-Shot)	61.01
HellaSwag (10-Shot)	81.55
MMLU (5-Shot)	58.22
TruthfulQA (0-shot)	60.40
Winogrande (5-shot)	76.09
GSM8k (5-shot)	31.08