metadata

license: apache-2.0
tags:
  - moe
  - frankenmoe
  - merge
  - mergekit
  - lazymergekit
  - phi3_mergekit
  - microsoft/Phi-3-mini-128k-instruct
base_model:
  - microsoft/Phi-3-mini-128k-instruct
  - microsoft/Phi-3-mini-128k-instruct

MixtureOfPhi3

MixtureOfPhi3 is a Mixure of Experts (MoE) made with the following models using mergekit:

This has been created using LazyMergekit-Phi3

This run is only for development purposes, since merging 2 identical models does not bring any performance benefits, but once specialized finetunes of Phi3 models will be available, it will be a starting point for creating MoE from them.

©️ Credits

mlabonne's phixtral where I adapted the inference code to Phi3's architecture.
mergekit code which I tweaked to merge Phi3s

These have been merged using cheap_embed where each model is assigned a vector representation of words - such as experts for scientific work, reasoning, math etc.

Try your own in the link above !

🧩 Configuration

base_model: microsoft/Phi-3-mini-128k-instruct
gate_mode: cheap_embed
dtype: float16
experts:
  - source_model: microsoft/Phi-3-mini-128k-instruct
    positive_prompts: ["research, logic, math, science"]
  - source_model: microsoft/Phi-3-mini-128k-instruct
    positive_prompts: ["creative, art"]

💻 Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer


model = "paulilioaica/MixtureOfPhi3"

tokenizer = AutoTokenizer.from_pretrained(model)
model = AutoModelForCausalLM.from_pretrained(
    model, 
    trust_remote_code=True, 
)

prompt="How many continents are there?"
input = f"<|system|>\nYou are a helpful AI assistant.<|end|>\n<|user|>{prompt}\n<|assistant|>"
tokenized_input = tokenizer.encode(input, return_tensors="pt")

outputs = model.generate(tokenized_input, max_new_tokens=128, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(tokenizer.decode(outputs[0]))