HelixNet-LMoE

HelixNet-LMoE is a simple LoRA based Mixture of Experts version of the HelixNet 3-model system by Migel Tissera.

Update : There is a 6bpw LMoE version that runs the entire 3-model system much faster, using 8 GB gpu mem in total. ExLlamaV2 version here: HelixNet-LMoE-6.0bpw-h6-exl2.

For each HelixNet model, a separate LoRA adapter was extracted :

• HelixNet-LMoE-Actor
• HelixNet-LMoE-Critic
• HelixNet-LMoE-Regenerator

These are then loaded together with the base Mistral 7b model to give the combined LMoE model.

As HelixNet processes its inputs using the actor, critic and regenerator actions, the corresponding LoRA adapter is dynamically enabled as required.

It is similar in approach to Airoboro's MoE implementation allowing GPU memory requirements in this (unquantized) instance to be reduced from 3 x 14GB to 1 x 14GB + 3 x 320MB. The LoRAs were extracted based on the process given in https://github.com/uukuguy/multi_loras, with a Rank of 64 and an Alpha of 128.

Prompt format:

SYSTEM: Elaborate on the topic using a Tree of Thoughts and backtrack when necessary to construct a clear, cohesive Chain of Thought reasoning. Always answer without hesitation.
USER: What is the relationship between Earth's atmosphere, magnetic field and gravity?
ASSISTANT:

Example Usage

The following is a code example on how to use HelixNet-LMoE. No special system-context messages are needed for the critic and the regenerator.
At the You: prompt, enter a question such as What is the relationship between Earth's atmosphere, magnetic field and gravity?

import torch, json
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

def load_model(model_path):
    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        torch_dtype=torch.float16,
        device_map="cuda",
        load_in_4bit=False,
        trust_remote_code=True,
    )
    return model

def load_tokenizer(model_path):
    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
    return tokenizer

def generate_text(instruction, adapter):
    # Select our required LoRA adapter
    adapter_model.set_adapter(adapter)
    
    tokens = base_tokenizer.encode(instruction)
    tokens = torch.LongTensor(tokens).unsqueeze(0)
    tokens = tokens.to("cuda")

    instance = {
        "input_ids": tokens,
        "top_p": 1.0,
        "temperature": 0.75,
        "generate_len": 1024,
        "top_k": 50,
    }

    length = len(tokens[0])
    with torch.no_grad():
        rest = adapter_model.generate(
            input_ids=tokens,
            max_length=length + instance["generate_len"],
            use_cache=True,
            do_sample=True,
            top_p=instance["top_p"],
            temperature=instance["temperature"],
            top_k=instance["top_k"],
            num_return_sequences=1,
            pad_token_id=base_tokenizer.eos_token_id,
        )
    output = rest[0][length:]
    string = base_tokenizer.decode(output, skip_special_tokens=True)
    return f"{string}"

# Load our base Mistral 7B model and tokenizer
base_model = load_model("mistralai/Mistral-7B-v0.1")
base_tokenizer = load_tokenizer("mistralai/Mistral-7B-v0.1")

# Load in our three different LoRA adapters for the actor, critic and regenerator
adapter_model = PeftModel.from_pretrained(base_model, "rhysjones/HelixNet-LMoE-Actor", "actor")
adapter_model.load_adapter("rhysjones/HelixNet-LMoE-Critic", adapter_name="critic")
adapter_model.load_adapter("rhysjones/HelixNet-LMoE-Regenerator", adapter_name="regenerator")

system_prompt = "You are HelixNet. Elaborate on the topic using a Tree of Thoughts and backtrack when necessary to construct a clear, cohesive Chain of Thought reasoning. Always answer without hesitation."

while True:
    user_input = input("You: ")

    prompt_actor = f"SYSTEM: {system_prompt} \nUSER: {user_input} \nASSISTANT: "
    actor_response = generate_text(prompt_actor, "actor")
    print(f"ACTOR: {actor_response}\n\n")

    prompt_critic = f"SYSTEM: {system_prompt} \nUSER: {user_input} \nRESPONSE: {actor_response} \nCRITIQUE:"
    critic_response = generate_text(prompt_critic, "critic")
    print(f"CRITIQUE: {critic_response}\n\n")
    
    prompt_regenerator = f"SYSTEM: {system_prompt} \nUSER: {user_input} \nRESPONSE: {actor_response} \nCRITIQUE: {critic_response} \nREGENERATOR:"
    regenerator_response = generate_text(prompt_regenerator, "regenerator")
    print(f"REGENERATION: {regenerator_response}")

LLM Evaluation

Evaluation on a merged version of each base+lora models has yet to be done on the HuggingFaceH4 Open LLM Leaderboard to see how it compares to the equivalent full HelixNet model.

HelixNet Details

HelixNet is a Deep Learning architecture consisting of 3 x Mistral-7B LLMs. It has an actor, a critic, and a regenerator. The actor LLM produces an initial response to a given system-context and a question. The critic then takes in as input, a tuple of (system-context, question, response) and provides a critique based on the provided answer to the given system-context and the question. Its job is not to criticize, but to provide an intelligent critique so that the answer can be modified/regenerated to address the question better. Finally, the regenerator takes in a tuple of (system-context, question, response, critique) and regenerates the answer.

HelixNet is insprired from an actor-critic architecture most prominent in Reinforcement Learning algorithms. The name derives from Helix, referring to the spiral structure of a DNA molecule. It symbolizes the intertwined nature of the three networks, working in tandem, much like the strands of a DNA molecule.

HelixNet regenerates very pleasing and accurate responses, due to the entropy preservation of the regenerator. The regenerator was only trained on a dataset of 1000 samples, similar to Meta's LIMA. The actor network here was trained on about 250K very high-quality samples, and the critic network was trained on further 10K samples.

Full details on how HelixNet was trained and evaluated is located at https://huggingface.co/migtissera/HelixNet