metadata

license: other
tags:
  - moe
  - merge
  - mergekit
  - Mistral
  - openchat/openchat-3.5-1210
  - beowolx/CodeNinja-1.0-OpenChat-7B
  - maywell/PiVoT-0.1-Starling-LM-RP
  - WizardLM/WizardMath-7B-V1.1
license_name: microsoft-research-license
license_link: https://huggingface.co/WizardLM/WizardMath-7B-V1.1/resolve/main/LICENSE
model-index:
  - name: Beyonder-4x7B-v2
    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: 68.77
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/Beyonder-4x7B-v2
          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: 86.8
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/Beyonder-4x7B-v2
          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: 65.1
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/Beyonder-4x7B-v2
          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: 60.68
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/Beyonder-4x7B-v2
          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: 80.9
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/Beyonder-4x7B-v2
          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: 71.72
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=mlabonne/Beyonder-4x7B-v2
          name: Open LLM Leaderboard

Beyonder-4x7B-v2

This model is a Mixture of Experts (MoE) made with mergekit (mixtral branch). It uses the following base models:

The recommended context length is 8k.

⚡ Quantized models

Thanks to TheBloke for the quantized models:

GGUF: https://huggingface.co/TheBloke/Beyonder-4x7B-v2-GGUF
AWQ: https://huggingface.co/TheBloke/Beyonder-4x7B-v2-AWQ
GPTQ: https://huggingface.co/TheBloke/Beyonder-4x7B-v2-GPTQ
EXL2: https://huggingface.co/bartowski/Beyonder-4x7B-v2-exl2

🏆 Evaluation

Beyonder-4x7B-v2 is competitive with Mixtral-8x7B-Instruct-v0.1 on the Open LLM Leaderboard, while only having 4 experts instead of 8.

It also displays a significant improvement over the individual experts.

It also performs very well compared to other models on Nous benchmark suite. It's almost as good as the best Yi-34B fine-tune, which is a much bigger model: 24.2B parameters + only two experts are selected during inference (so ~12B) vs. 34B param.

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
Beyonder-4x7B-v2	45.29	75.95	60.86	46.4	57.13
NeuralHermes-2.5-Mistral-7B	43.67	73.24	55.37	41.76	53.51
OpenHermes-2.5-Mistral-7B	42.75	72.99	52.99	40.94	52.42
Nous-Hermes-2-SOLAR-10.7B	47.79	74.69	55.92	44.84	55.81
Nous-Hermes-2-Yi-34B	50.27	76.00	60.34	46.69	58.33

AGIEval

Task	Version	Metric	Value		Stderr
agieval_aqua_rat	0	acc	23.62	±	2.67
		acc_norm	23.62	±	2.67
agieval_logiqa_en	0	acc	41.47	±	1.93
		acc_norm	43.01	±	1.94
agieval_lsat_ar	0	acc	23.04	±	2.78
		acc_norm	23.48	±	2.80
agieval_lsat_lr	0	acc	51.57	±	2.22
		acc_norm	52.94	±	2.21
agieval_lsat_rc	0	acc	64.31	±	2.93
		acc_norm	64.68	±	2.92
agieval_sat_en	0	acc	79.13	±	2.84
		acc_norm	79.13	±	2.84
agieval_sat_en_without_passage	0	acc	43.20	±	3.46
		acc_norm	43.20	±	3.46
agieval_sat_math	0	acc	34.55	±	3.21
		acc_norm	32.27	±	3.16

GPT4All

Task	Version	Metric	Value		Stderr
arc_challenge	0	acc	61.86	±	1.42
		acc_norm	64.51	±	1.40
arc_easy	0	acc	85.06	±	0.73
		acc_norm	82.45	±	0.78
boolq	1	acc	88.35	±	0.56
hellaswag	0	acc	68.04	±	0.47
		acc_norm	85.12	±	0.36
openbookqa	0	acc	37.80	±	2.17
		acc_norm	48.60	±	2.24
piqa	0	acc	83.08	±	0.87
		acc_norm	83.95	±	0.86
winogrande	0	acc	78.69	±	1.15

TruthfulQA

Task	Version	Metric	Value		Stderr
truthfulqa_mc	1	mc1	44.55	±	1.74
		mc2	60.86	±	1.57

Bigbench

Task	Version	Metric	Value		Stderr
bigbench_causal_judgement	0	multiple_choice_grade	58.95	±	3.58
bigbench_date_understanding	0	multiple_choice_grade	66.40	±	2.46
bigbench_disambiguation_qa	0	multiple_choice_grade	48.84	±	3.12
bigbench_geometric_shapes	0	multiple_choice_grade	22.56	±	2.21
		exact_str_match	13.37	±	1.80
bigbench_logical_deduction_five_objects	0	multiple_choice_grade	30.40	±	2.06
bigbench_logical_deduction_seven_objects	0	multiple_choice_grade	20.57	±	1.53
bigbench_logical_deduction_three_objects	0	multiple_choice_grade	52.00	±	2.89
bigbench_movie_recommendation	0	multiple_choice_grade	44.40	±	2.22
bigbench_navigate	0	multiple_choice_grade	52.10	±	1.58
bigbench_reasoning_about_colored_objects	0	multiple_choice_grade	69.75	±	1.03
bigbench_ruin_names	0	multiple_choice_grade	55.36	±	2.35
bigbench_salient_translation_error_detection	0	multiple_choice_grade	23.65	±	1.35
bigbench_snarks	0	multiple_choice_grade	77.35	±	3.12
bigbench_sports_understanding	0	multiple_choice_grade	73.02	±	1.41
bigbench_temporal_sequences	0	multiple_choice_grade	46.80	±	1.58
bigbench_tracking_shuffled_objects_five_objects	0	multiple_choice_grade	22.08	±	1.17
bigbench_tracking_shuffled_objects_seven_objects	0	multiple_choice_grade	19.03	±	0.94
bigbench_tracking_shuffled_objects_three_objects	0	multiple_choice_grade	52.00	±	2.89

🧩 Configuration

base_model: mlabonne/Marcoro14-7B-slerp
experts:
  - source_model: openchat/openchat-3.5-1210
    positive_prompts:
    - "chat"
    - "assistant"
    - "tell me"
    - "explain"
  - source_model: beowolx/CodeNinja-1.0-OpenChat-7B
    positive_prompts:
    - "code"
    - "python"
    - "javascript"
    - "programming"
    - "algorithm"
  - source_model: maywell/PiVoT-0.1-Starling-LM-RP
    positive_prompts:
    - "storywriting"
    - "write"
    - "scene"
    - "story"
    - "character"
  - source_model: WizardLM/WizardMath-7B-V1.1
    positive_prompts:
    - "reason"
    - "math"
    - "mathematics"
    - "solve"
    - "count"

💻 Usage

Here's a notebook to run this model in 4-bit precision using a free T4 GPU on Google Colab.

!pip install -qU transformers bitsandbytes accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "mlabonne/Beyonder-4x7B-v2"

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"])

Output:

A Mixture of Experts (ME) is a machine learning technique that combines multiple expert models to make predictions or decisions. Each expert model is specialized in a different aspect of the problem, and their outputs are combined to produce a more accurate and robust solution. This approach allows the model to leverage the strengths of individual experts and compensate for their weaknesses, improving overall performance.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	72.33
AI2 Reasoning Challenge (25-Shot)	68.77
HellaSwag (10-Shot)	86.80
MMLU (5-Shot)	65.10
TruthfulQA (0-shot)	60.68
Winogrande (5-shot)	80.90
GSM8k (5-shot)	71.72