metadata

language:
  - en
license: cc-by-nc-nd-4.0
library_name: transformers
tags:
  - moe
  - merge
  - medical
  - mergekit
datasets:
  - medmcqa
  - cognitivecomputations/samantha-data
  - jondurbin/bagel-v0.3
base_model:
  - sethuiyer/Dr_Samantha_7b_mistral
  - fblgit/UNA-TheBeagle-7b-v1
pipeline_tag: text-generation
model-index:
  - name: MedleyMD
    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: 66.47
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=sethuiyer/MedleyMD
          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: 86.06
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=sethuiyer/MedleyMD
          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=sethuiyer/MedleyMD
          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: 52.46
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=sethuiyer/MedleyMD
          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: 80.27
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=sethuiyer/MedleyMD
          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: 68.99
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=sethuiyer/MedleyMD
          name: Open LLM Leaderboard

MedleyMD

MedleyMD is a Mixure of Experts (MoE) made with the following models using LazyMergekit:

These models were chosen because fblgit/UNA-TheBeagle-7b-v1 has excellent performance for a 7B parameter model and Dr.Samantha has capabilities of a medical knowledge-focused model (trained on USMLE databases and doctor-patient interactions) with the philosophical, psychological, and relational understanding, scoring 68.82% in topics related to clinical domain and psychology.

Benchmark

On a synthetic benchmark of 35 medical diagnosis questions generated by GPT-4, GPT-4 also being an evaluator, MedleyMD scored 96.25/100.

Nous Benchmark numbers shall be available soon.

🧩 Configuration

base_model: OpenPipe/mistral-ft-optimized-1227
gate_mode: hidden
dtype: bfloat16

experts:
  - source_model: sethuiyer/Dr_Samantha_7b_mistral
    positive_prompts: ["differential diagnosis", "Clinical Knowledge", "Medical Genetics", "Human Aging", "Human Sexuality", "College Medicine", "Anatomy", "College Biology", "High School Biology", "Professional Medicine", "Nutrition", "High School Psychology", "Professional Psychology", "Virology"]

  - source_model: fblgit/UNA-TheBeagle-7b-v1
    positive_prompts: ["How do you", "Explain the concept of", "Give an overview of", "Compare and contrast between", "Provide information about", "Help me understand", "Summarize", "Make a recommendation on", "chat", "math", "reason", "mathematics", "solve", "count", "python", "javascript", "programming", "algorithm", "tell me", "assistant"]

GGUF

medleymd.Q4_K_M [7.2GB]
medleymd.Q5_K_M [9.13GB]

Ollama

MedleyMD can be used in ollama by runningollama run stuehieyr/medleymd in your terminal.

If you have limited computing resources, check out this video to learn how to run it on a Google Colab backend.

Prompt format:

This model uses ChatML prompt format.

<|im_start|>system
You are Medley, a helpful AI assistant.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

💻 Usage

!pip install -qU transformers bitsandbytes accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "sethuiyer/MedleyMD"

tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    model_kwargs={"torch_dtype": torch.bfloat16, "load_in_4bit": True},
)

generation_kwargs = {
    "max_new_tokens": 512,
    "do_sample": True,
    "temperature": 0.7,
    "top_k": 50,
    "top_p": 95,
}

messages = [{"role":"system", "content":"You are an helpful AI assistant. Please use </s> when you want to end the answer."},
{"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, **generation_kwargs)
print(outputs[0]["generated_text"])

A Mixture of Experts (Mixout) is a neural network architecture that combines the strengths of multiple expert networks to make a more accurate and robust prediction.
It is composed of a topmost gating network that assigns weights to each expert network based on their performance on past input samples.
The expert networks are trained independently, and the gating network learns to choose the best combination of these experts to make the final prediction.
Mixout demonstrates a stronger ability to handle complex data distributions and is more efficient in terms of training time and memory usage compared to a
traditional ensemble approach.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	69.89
AI2 Reasoning Challenge (25-Shot)	66.47
HellaSwag (10-Shot)	86.06
MMLU (5-Shot)	65.10
TruthfulQA (0-shot)	52.46
Winogrande (5-shot)	80.27
GSM8k (5-shot)	68.99