README.md

metadata

language:
  - en
license: apache-2.0
tags:
  - merge
datasets:
  - Locutusque/inst_mix_v2_top_100k
pipeline_tag: text-generation
widget:
  - text: >-
      <|USER|> Design a Neo4j database and Cypher function snippet to Display
      Extreme Dental hygiene: Using Mouthwash for Analysis for Beginners.
      Implement if/else or switch/case statements to handle different conditions
      related to the Consent. Provide detailed comments explaining your control
      flow and the reasoning behind each decision. <|ASSISTANT|> 
  - text: '<|USER|> Write me a story about a magical place. <|ASSISTANT|> '
  - text: >-
      <|USER|> Write me an essay about the life of George Washington
      <|ASSISTANT|> 
  - text: '<|USER|> Solve the following equation 2x + 10 = 20 <|ASSISTANT|> '
  - text: >-
      <|USER|> Craft me a list of some nice places to visit around the world.
      <|ASSISTANT|> 
  - text: >-
      <|USER|> How to manage a lazy employee: Address the employee verbally.
      Don't allow an employee's laziness or lack of enthusiasm to become a
      recurring issue. Tell the employee you're hoping to speak with them about
      workplace expectations and performance, and schedule a time to sit down
      together. Question: To manage a lazy employee, it is suggested to talk to
      the employee. True, False, or Neither? <|ASSISTANT|> 
inference:
  parameters:
    temperature: 0.5
    do_sample: true
    top_p: 0.5
    top_k: 30
    max_new_tokens: 250
    repetition_penalty: 1.15
model-index:
  - name: LocutusqueXFelladrin-TinyMistral248M-Instruct
    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: 24.74
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Locutusque/LocutusqueXFelladrin-TinyMistral248M-Instruct
          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: 27.79
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Locutusque/LocutusqueXFelladrin-TinyMistral248M-Instruct
          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: 26.12
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Locutusque/LocutusqueXFelladrin-TinyMistral248M-Instruct
          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: 40.12
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Locutusque/LocutusqueXFelladrin-TinyMistral248M-Instruct
          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: 49.09
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Locutusque/LocutusqueXFelladrin-TinyMistral248M-Instruct
          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: 0
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Locutusque/LocutusqueXFelladrin-TinyMistral248M-Instruct
          name: Open LLM Leaderboard

LocutusqueXFelladrin-TinyMistral248M-Instruct

This model was created by merging Locutusque/TinyMistral-248M-Instruct and Felladrin/TinyMistral-248M-SFT-v4 using mergekit. After the two models were merged, the resulting model was further trained on ~20,000 examples on the Locutusque/inst_mix_v2_top_100k at a low learning rate to further normalize weights. The following is the YAML config used to merge:

models:
  - model: Felladrin/TinyMistral-248M-SFT-v4
    parameters:
      weight: 0.5
  - model: Locutusque/TinyMistral-248M-Instruct
    parameters:
      weight: 1.0
merge_method: linear
dtype: float16

The resulting model combines the best of both worlds. With Locutusque/TinyMistral-248M-Instruct's coding capabilities and reasoning skills, and Felladrin/TinyMistral-248M-SFT-v4's low hallucination and instruction-following capabilities. The resulting model has an incredible performance considering its size.

Evaluation

Found in the Open LLM Leaderboard.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	27.98
AI2 Reasoning Challenge (25-Shot)	24.74
HellaSwag (10-Shot)	27.79
MMLU (5-Shot)	26.12
TruthfulQA (0-shot)	40.12
Winogrande (5-shot)	49.09
GSM8k (5-shot)	0.00