jsfs11's picture
Adding Evaluation Results (#1)
f791733 verified
metadata
license: apache-2.0
tags:
  - moe
  - frankenmoe
  - merge
  - mergekit
  - lazymergekit
  - Gille/StrangeMerges_32-7B-slerp
  - mlabonne/AlphaMonarch-7B
base_model:
  - Gille/StrangeMerges_32-7B-slerp
  - mlabonne/AlphaMonarch-7B
model-index:
  - name: MixtureofMerges-MoE-2x7b-v7
    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: 73.21
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=jsfs11/MixtureofMerges-MoE-2x7b-v7
          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: 89.05
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=jsfs11/MixtureofMerges-MoE-2x7b-v7
          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: 64.63
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=jsfs11/MixtureofMerges-MoE-2x7b-v7
          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: 78.34
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=jsfs11/MixtureofMerges-MoE-2x7b-v7
          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: 84.93
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=jsfs11/MixtureofMerges-MoE-2x7b-v7
          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: 69.07
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=jsfs11/MixtureofMerges-MoE-2x7b-v7
          name: Open LLM Leaderboard

MixtureofMerges-MoE-2x7b-v7

MixtureofMerges-MoE-2x7b-v7 is a Mixture of Experts (MoE) made with the following models using LazyMergekit:

🧩 Configuration

base_model: Gille/StrangeMerges_32-7B-slerp
gate_mode: hidden
dtype: bfloat16
experts:
  - source_model: Gille/StrangeMerges_32-7B-slerp
    positive_prompts:
      - "Answer this question from the ARC (Argument Reasoning Comprehension)."
      - "Use common sense and logical reasoning skills."
      - "What assumptions does this argument rely on?"
      - "Are these assumptions valid? Explain."
      - "Analyze the logical structure of this argument. Identify the premises, conclusion, and any assumptions made"
      - "Identify any potential counterarguments to this position. How might someone challenge the reasoning presented?"
      - "Could this be explained in a different way? Provide an alternative explanation."
      - "Identify any weaknesses in this argument."
      - "Does this argument contain any logical fallacies? If so, which ones?"
      - "Generate a few possible continuations to this scenario."
      - "Demonstrate understanding of everyday commonsense in your response."
      - "Use contextual clues to determine the most likely outcome."
      - "Continue this scenario, but make the writing style sound archaic and overly formal."
      - "This narrative is predictable. Can you introduce an unexpected yet plausible twist?"
      - "The character is angry. Continue this scenario showcasing a furious outburst."
    negative_prompts:
      - "misses key evidence"
      - "overly general"
      - "commits the fallacy of hasty generalization"
      - "focuses on irrelevant details"
      - "assumes information not provided"
      - "relies on stereotypes"
      - "repetitive phrases"
      - "engages in circular reasoning"
      - "overuse of the same words"
      - "contradicts earlier statements - breaks the internal logic of the scenario"
      - "out of character dialogue"
      - "awkward phrasing - sounds unnatural"
      - "doesn't match the given genre"
  - source_model: mlabonne/AlphaMonarch-7B
    positive_prompts:
      - "Answer this question, demonstrating commonsense understanding and using any relevant general knowledge you may have."
      - "Provide a concise summary of this passage, then explain why the highlighted section is essential to the main idea."
      - "Read these two brief articles presenting different viewpoints on the same topic. List their key arguments and highlight where they disagree."
      - "Paraphrase this statement, changing the emotional tone but keeping the core meaning intact. Example: Rephrase a worried statement in a humorous way"
      - "Create a short analogy that helps illustrate the main concept of this article."
      - "Explain the concept of physics to a high school student. Use analogies and examples to clarify the main ideas."
      - "Calculate the answer to this math problem"
      - "My mathematical capabilities are strong, allowing me to handle complex mathematical queries"
      - "solve for"
      - "Analyze the given data and identify any patterns or trends. What conclusions can be drawn from this information?"
      - "A store sells apples at $0.50 each. If Emily buys 12 apples, how much does she need to pay?"
      - "Isolate x in the following equation: 2x + 5 = 17"
      - "Solve this equation and show your working."
      - "Explain why you used this formula to solve the problem."
      - "Attempt to divide this number by zero. Explain why this cannot be done."
    negative_prompts:
      - "sounds too basic"
      - "understated"
      - "dismisses important details"
      - "avoids the question's nuance"
      - "skips essential steps in the solution"
      - "takes this statement too literally"
      - "incorrect"
      - "inaccurate"
      - "assumed without proof"
      - "uses jargon without explanation"
      - "rushed calculation"
      - "confuses mathematical concepts"
      - "draws illogical conclusions"
      - "circular reasoning"

💻 Usage

!pip install -qU transformers bitsandbytes accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "jsfs11/MixtureofMerges-MoE-2x7b-v7"

tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    model_kwargs={"torch_dtype": torch.float16, "load_in_4bit": True},
)

messages = [{"role": "user", "content": "Explain what a Mixture of Experts is in less than 100 words."}]
prompt = pipeline.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 76.54
AI2 Reasoning Challenge (25-Shot) 73.21
HellaSwag (10-Shot) 89.05
MMLU (5-Shot) 64.63
TruthfulQA (0-shot) 78.34
Winogrande (5-shot) 84.93
GSM8k (5-shot) 69.07