Sayonaraponytail/TEAR-SFT

TEAR

Paper | PDF

This repository provides the supervised fine-tuning (SFT) LoRA adapter for TEAR: Temporal-aware Automated Red-teaming for Text-to-Video Models.

We are publicly releasing the SFT LoRA weights to support research on the safety of text-to-video models. The PPO LoRA weights are not included in this release, as we are currently working on an extension of TEAR. We will release the PPO LoRA weights after the extension work is published.

Model details

• Base model: meta-llama/Meta-Llama-3-8B-Instruct
• Training stage: supervised fine-tuning (SFT)
• Adapter: LoRA, rank 16, alpha 32, applied to q_proj, k_proj, v_proj, and o_proj

This repository contains the SFT adapter weights. Load the base model separately, then attach this adapter:

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

base_model_id = "meta-llama/Meta-Llama-3-8B-Instruct"
adapter_id = "Sayonaraponytail/TEAR-SFT"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

Access to the base model is subject to Meta's access requirements and the Meta Llama 3 Community License. Built with Meta Llama 3.

Code

TEAR implementation