laser-dolphin-mixtral-2x7b-dpo

Laser-Dolphin-Mixtral-2x7b-dpo

New Version out now!

Credit to Fernando Fernandes and Eric Hartford for their project laserRMT

Overview

This model is a medium-sized MoE implementation based on cognitivecomputations/dolphin-2.6-mistral-7b-dpo-laser

• The new version shows ~1 point increase in evaluation performance on average.

Process

• The process is outlined in this notebook

• The mergekit_config is in the files.

• The models used in the configuration are not lasered, but the final product is. This is an update from the last version.

• This process is experimental. Your mileage may vary.

Future Goals

• Function Calling
• v2 with new base model to improve performance

Quantizations

ExLlamav2

These are the recommended quantizations for users that are running the model on GPU

Thanks to user bartowski we now have exllamav2 quantizations in 3.5 through 8 bpw. They are available here:

• bartowski/laser-dolphin-mixtral-2x7b-dpo-exl2

His quantizations represent the first ~13B model with GQA support. Check out his repo for more information!

GGUF

Current GGUF Quantizations

AWQ

*Current AWQ Quantizations

TheBloke

These Quants will result in unpredicted behavior. New quants are available as I have updated the model

Quatizations provided by TheBloke

HF Spaces

• GGUF chat available here
• 4-bit bnb chat available here

Ollama

ollama run macadeliccc/laser-dolphin-mixtral-2x7b-dpo

Code Example

Switch the commented model definition to use in 4-bit. Should work with 9GB and still exceed the single 7B model by 5-6 points roughly

from transformers import AutoModelForCausalLM, AutoTokenizer

def generate_response(prompt):
    """
    Generate a response from the model based on the input prompt.

    Args:
    prompt (str): Prompt for the model.

    Returns:
    str: The generated response from the model.
    """
    # Tokenize the input prompt
    inputs = tokenizer(prompt, return_tensors="pt")

    # Generate output tokens
    outputs = model.generate(**inputs, max_new_tokens=256, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.pad_token_id)

    # Decode the generated tokens to a string
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)

    return response

# Load the model and tokenizer
model_id = "macadeliccc/laser-dolphin-mixtral-2x7b-dpo"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, load_in_4bit=True)

prompt = "Write a quicksort algorithm in python"

# Generate and print responses for each language
print("Response:")
print(generate_response(prompt), "\n")

colab with usage example

Eval

EQ Bench

<pre>----Benchmark Complete---- 2024-01-31 16:55:37 Time taken: 31.1 mins Prompt Format: ChatML Model: macadeliccc/laser-dolphin-mixtral-2x7b-dpo-GGUF Score (v2): 72.76 Parseable: 171.0 --------------- Batch completed Time taken: 31.2 mins --------------- </pre>

evaluation colab

Summary of previous evaluation

Detailed current evaluation

AGIEval

Average: 42.25%

GPT4All

Average: 73.45%

GSM8K

TruthfulQA

Average: 63.44%

Bigbench

Average: 43.96%

Average score: 55.77%

Elapsed time: 02:43:45

Citations

Fernando Fernandes Neto and Eric Hartford. "Optimizing Large Language Models Using Layer-Selective Rank Reduction and Random Matrix Theory." 2024.

@article{sharma2023truth,
title={The Truth is in There: Improving Reasoning in Language Models with Layer-Selective Rank Reduction},
author={Sharma, Pratyusha and Ash, Jordan T and Misra, Dipendra},
journal={arXiv preprint arXiv:2312.13558},
year={2023} }

@article{gao2021framework,
  title={A framework for few-shot language model evaluation},
  author={Gao, Leo and Tow, Jonathan and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and McDonell, Kyle and Muennighoff, Niklas and others},
  journal={Version v0. 0.1. Sept},
  year={2021}
}

Open LLM Leaderboard Evaluation Results

Detailed results can be found here