README.md

metadata

language:
  - en
license: apache-2.0
datasets:
  - HuggingFaceTB/cosmopedia
  - EleutherAI/proof-pile-2
  - bigcode/the-stack-dedup
  - math-ai/AutoMathText
metrics:
  - accuracy
  - code_eval
model-index:
  - name: Mistral_Pro_8B_v0.1
    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: 62.2
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=TencentARC/Mistral_Pro_8B_v0.1
          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: 82.13
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=TencentARC/Mistral_Pro_8B_v0.1
          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: 61.74
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=TencentARC/Mistral_Pro_8B_v0.1
          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: 49.32
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=TencentARC/Mistral_Pro_8B_v0.1
          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: 76.8
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=TencentARC/Mistral_Pro_8B_v0.1
          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: 34.19
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=TencentARC/Mistral_Pro_8B_v0.1
          name: Open LLM Leaderboard

Mistral-Pro-8B Model Card

Model Description

Mistral-Pro is a progressive version of the original Mistral model, enhanced by the addition of Transformer blocks. It specializes in integrating both general language understanding and domain-specific knowledge, particularly in programming and mathematics.

Development and Training

Developed by Tencent's ARC Lab, Mistral-Pro is an 8 billion parameter model. It's an expansion of Mistral-7B, further trained on code and math corpora.

Intended Use

This model is designed for a wide range of NLP tasks, with a focus on programming, mathematics, and general language tasks. It suits scenarios requiring integration of natural and programming languages.

Performance

Mistral_Pro_8B_v0.1 showcases superior performance on a range of benchmarks. It enhances the code and math performance of Mistral. Furthermore, it matches the performance of the recently dominant model, Gemma.

Overall Performance on Languages, math and code tasks

Model	ARC	Hellaswag	MMLU	TruthfulQA	Winogrande	GSM8K	HumanEval
Gemma-7B	61.9	82.2	64.6	44.8	79.0	50.9	32.3
Mistral-7B	60.8	83.3	62.7	42.6	78.0	39.2	28.7
Mistral_Pro_8B_v0.1	63.2	82.6	60.6	48.3	78.9	50.6	32.9

Limitations

While Mistral-Pro addresses some limitations of previous models in the series, it may still encounter challenges specific to highly specialized domains or tasks.

Ethical Considerations

Users should be aware of potential biases in the model and use it responsibly, considering its impact on various applications.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	61.06
AI2 Reasoning Challenge (25-Shot)	62.20
HellaSwag (10-Shot)	82.13
MMLU (5-Shot)	61.74
TruthfulQA (0-shot)	49.32
Winogrande (5-shot)	76.80
GSM8k (5-shot)	34.19