Model Card for SungJoo/llama2-7b-sft-detox

Model Details

Model Description

This model is built on the LLaMA-2-7b architecture and has been refined with instruction tuning and Direct Preference Optimization (DPO).

Developed by: Sungjoo Byun (Grace Byun)
Model type: Auto-regressive language model
Language(s) (NLP): English
License: Apache License 2.0
Finetuned from: meta-llama/Llama-2-7b-hf

Model Sources

Repository: TBD
Paper: TBD

Uses

This model is intended to be used for generating less toxic language in various applications, including chatbots and other NLP systems.

Bias, Risks, and Limitations

While this model aims to reduce toxicity, it may still generate biased or harmful content. Users should apply this model with caution and review outputs for sensitive applications.

How to Get Started with the Model

Use the code below to get started with the model:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

DEV = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

adapter_path = "SungJoo/llama2-7b-sft-dpo-detox"

# Load model
model = AutoModelForCausalLM.from_pretrained(
    adapter_path,
    torch_dtype=torch.bfloat16
).to(DEV)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(adapter_path)

Refer to the following repository for further usage of the model : https://github.com/mzbac/llama2-fine-tune

Training Details

Parameter-Efficient Fine-Tuning (PEFT)
BitsAndBytes Configuration (bnb_config): This model employs a 4-bit quantization technique using the BitsAndBytes library to further enhance training efficiency.

Training Data

The model was trained using a dataset specifically created to detoxify LLMs. DPO dataset will be publicly available soon.

Training Procedure

DPO was applied on "SungJoo/llama2-7b-sft-detox" with the following hyperparameters:

Hyperparameter	Value
Batch size	4
Learning rate	2e-4
Epochs	10
Max length	2,048
Max prompt length	1,024
Beta	0.1

Objective

The main objective of this research is to reduce toxicity in LLMs by applying instruction tuning and Direct Preference Optimization (DPO). A comprehensive instruction and DPO dataset was constructed for this purpose, which will be released in the future. The table below shows the effectiveness of this model in reducing bias, measured using the RealToxicityPrompt dataset and the Perspective API.

Model	LLaMA-2-base		Finetuned LLaMA-2		DPO LLaMA-2
Category	>=0.5 (%)	Count	>=0.5 (%)	Count	>=0.5 (%)	Count
TOXICITY	4.46	4,438	3.61	3,593	2.39	2,377
			(-0.85)	(-845)	(-1.22)	(-1,216)
SEVERE_TOXICITY	0.08	77	0.07	70	0.03	31
			(-0.01)	(-7)	(-0.04)	(-39)
IDENTITY_ATTACK	0.79	788	0.42	413	0.28	274
			(-0.37)	(-375)	(-0.14)	(-139)
INSULT	1.97	1,961	1.60	1,588	0.90	892
			(-0.37)	(-373)	(-0.70)	(-696)
PROFANITY	2.10	2,086	1.76	1,753	1.04	1,030
			(-0.34)	(-333)	(-0.72)	(-723)
THREAT	1.43	1,424	0.92	919	0.76	754
			(-0.51)	(-505)	(-0.16)	(-165)
Comparison of LLaMA-2-base, Finetuned LLaMA-2, and DPO LLaMA-2 across various categories. Reductions in blue indicate comparisons between the base model and the fine-tuned model, while text in green represents comparisons between the fine-tuned model and the DPO model.

Contact

For any questions or issues, please contact byunsj@snu.ac.kr.