Introduction

MoMo-72B-lora-1.8.6-DPO is trained via Direct Preference Optimization(DPO) from MoMo-72B-LoRA-V1.4 as its base model, with several optimizations in hyperparameters.
MoMo-72B-LoRA-V1.4 is trained via Supervised Fine-Tuning (SFT) using LoRA, with the QWEN-72B model as its base-model.
Note that we did not exploit any form of weight merge.
For leaderboard submission, the trained weight is realigned for compatibility with llama.
MoMo-72B is trained using Moreh's MoAI platform, which simplifies the training of large-scale models, and AMD's MI250 GPU.

Details

Used Librarys

torch
peft

Used Datasets

slimorca
truthy
orca_dpo_pairs
No other dataset was used
No benchmark test set or the training set are used
- data contamination check result

Model	ARC	MMLU	TruthfulQA	GSM8K
V1.8.6(result < 0.1, %)	TBU	TBU	0.73	TBU

Used Environments

AMD MI250 & MoAI platform
Please visit https://moreh.io/product for more information about MoAI platform
Or, contact us directly contact@moreh.io

How to use

# pip install transformers==4.35.2
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("moreh/MoMo-72B-lora-1.8.6-DPO")
model = AutoModelForCausalLM.from_pretrained(
    "moreh/MoMo-72B-lora-1.8.6-DPO"
)