README.md

metadata

language:
  - zh
  - ja
  - ko
  - tw
license: apache-2.0
tags:
  - moe
  - merge
  - mergekit
  - lazymergekit
  - MediaTek-Research/Breeze-7B-Instruct-v0.1
  - augmxnt/shisa-7b-v1
  - beomi/OPEN-SOLAR-KO-10.7B
model-index:
  - name: EastAsia-4x7B-Moe-experiment
    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: 39.51
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Heng666/EastAsia-4x7B-Moe-experiment
          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: 48.92
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Heng666/EastAsia-4x7B-Moe-experiment
          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: 56.2
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Heng666/EastAsia-4x7B-Moe-experiment
          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: 49.83
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Heng666/EastAsia-4x7B-Moe-experiment
          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: 58.09
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Heng666/EastAsia-4x7B-Moe-experiment
          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: 0.15
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=Heng666/EastAsia-4x7B-Moe-experiment
          name: Open LLM Leaderboard

EastAsia-4x7B-Moe-experiment

EastAsia-4x7B-Moe-experiment is a Mixure of Experts (MoE) made with the following models using LazyMergekit:

• MediaTek-Research/Breeze-7B-Instruct-v0.1
• augmxnt/shisa-7b-v1
• beomi/OPEN-SOLAR-KO-10.7B

🧩 Configuration

gate_mode: hidden
dtype: bfloat16
base_model: mlabonne/Marcoro14-7B-slerp
experts:
  - source_model: MediaTek-Research/Breeze-7B-Instruct-v0.1
    positive_prompts:
    - "翻譯"
  - source_model: augmxnt/shisa-7b-v1
    positive_prompts:
    - "翻訳"
  - source_model: beomi/OPEN-SOLAR-KO-10.7B
    positive_prompts:
    - "번역"

💻 Usage

!pip install -qU transformers bitsandbytes accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "Heng666/EastAsia-4x7B-Moe-experiment"

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.	42.12
AI2 Reasoning Challenge (25-Shot)	39.51
HellaSwag (10-Shot)	48.92
MMLU (5-Shot)	56.20
TruthfulQA (0-shot)	49.83
Winogrande (5-shot)	58.09
GSM8k (5-shot)	0.15