Modelpublic

abbatea/Tutorbot-variation-DPO-Llama

sourceHugging Faceupdated 1y agoView on Hugging Face

0likes6downloads

Model Card

Model Card for Model ID

Model Card for LlamaDPOlora

This model is a fine-tuned version of meta-llama/Llama-3.1-8B-Instruct. It has been trained using DPO. The dataset it was build upon is a combination on [MathDial dataset](https://huggingface.co/datasets/eth-nlped/mathdial-chat/viewer/default/train?views%5B%5D=train&row=0) and generated model responses using MathDial as an input.

The model is optimized for:

—Conversational math problem solving
—Step-by-step reasoning in dialogue form
—Scaffolding

Repository: [Github code DPO Training and Datasets](https://github.com/abbatea/MathDial-SFT-and-DPO/tree/main/DPO_Finetuning)

Intended Use

This model is intended for use in:

—Interactive math tutoring
—Research in dialogue-based problem solving
—Educational tools

Example Usage

python


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

# Base model
basemodelname = "meta-llama/Llama-3.1-8B-Instruct"
base_model = AutoModelForCausalLM.from_pretrained(
    basemodelname,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load adapter on top
peft_model_path = "abbatea/Tutorbot-variation-DPO-Llama"
model = PeftModel.from_pretrained(base_model, peft_model_path)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(basemodelname)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.pad_token_id = tokenizer.eos_token_id



messages = [
    {"role": "user", "content": "Can you help me solve 3x + 5 = 20?"}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Citations

DPO:

bibtex

@inproceedings{rafailov2023direct,
    title        = {{Direct Preference Optimization: Your Language Model is Secretly a Reward Model}},
    editor       = {Alice Oh and Tristan Naumann and Amir Globerson and Kate Saenko and Moritz Hardt and Sergey Levine},
}