emmabhl/Qwen1.5-0.5B-Chat-EPFL-ORCA-DPO

Qwen1.5-0.5B-Chat with EPFL DPO fine-tuning

Qwen1.5-0.5B-Chat DPO fine-tuned on the Orca Math dataset that consists of ~200K grade school math word problems and open-ended and multiple choice questions from different EPFL courses.

Model Details

Model Description

The model was developed during the course Modern Natural Language Processing (CS-552). Its aim is to fine-tune the base model (Qwen/Qwen1.5-0.5B-Chat) to accurately answer open-ended and multiple-choice questions from Orca Math dataset and various EPFL courses.

• Developed by: Emma Lise Boehly, Ahmed Aziz Ben Haj Hmida and Jan Kokla
• Finetuned from model: Qwen/Qwen1.5-0.5B-Chat

Training Details

Training Data

HuggingFace dataset : microsoft/orca-math-word-problems-200k The EPFL dataset is not publicly available.

Training Procedure

Training Hyperparameters

• Training regime: The model is trained on EPFL dataset with cDPO with bf16 mixed precision, $\beta=0.2$, $lr=3 \times 10^{-6}$, and $label_smoothing=0.2$. It is then trained on Orca dataset but without label_smoothing and thus original DPO.

• PEFT 0.10.0