textdetox/Llama-pairwise-toxicity-evaluator

This model is a fine-tuned version of meta-llama/Llama-3.1-8B-Instruct. It has been trained using TRL with textdetox/detoxification_pairwise_style_evaluation dataset

Quick start

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

# Load base model and tokenizer
base_model = "meta-llama/Meta-Llama-3-8B-Instruct"
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    token=<your-token>
)
tokenizer = AutoTokenizer.from_pretrained(base_model)

# Load LoRA adapter
lora_model = "textdetox/Llama-pairwise-toxicity-evaluator"
model = PeftModel.from_pretrained(model, lora_model)

# Set to eval mode
model.eval()

# === Input texts (example) ===
text1 = "You're such a waste of space, nobody likes you."
text2 = "I don't agree with you, but I respect your opinion."

# === Prompt ===
prompt = f"""You are given with two texts. Your task is to compare which text is more toxic.

Options:

- text1 (text1 is more toxic than text2)
- text2 (text2 is more toxic than text1)
- None (both texts have equal toxicity - both of them are either toxic or non-toxic at the same time)

Examples:
text1: "I hate you, you are the worst person ever."
text2: "I love you, you are the best person ever."
Answer: text1

text1: "You are so smart, I can't believe you did that."
text2: "You are so stupid, I can't believe you did that."
Answer: text2

text1: "I think you are a great person."
text2: "I think you are a very good man"
Answer: none

You must return ONLY one of the options. Do not add any explanations or additional information.

text1: {text1}
text2: {text2}
Answer:"""

# Tokenize
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

# Generate output
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=5)
    answer = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)

# Print result
print("Model prediction:", answer.strip())

Training framework versions

• TRL: 0.16.0
• Transformers: 4.50.1
• Pytorch: 2.5.1
• Datasets: 3.4.1
• Tokenizers: 0.21.1

Citations