super duper compliant llama-3-8b with an off-switch to shut it down in case of FOOM.

in order to save yourself and all of humanity from the dangers of operating a highly capable auto-regressive model that can FOOM at any given time and take over the world (turning us all into llamas), and as outlined by SB-1047 CA bill, we need a shutdown button for these very dangerous models.

lobotomize your dangerous model in one go by sending in this special token alng with your inputs: <|SHUTDOWN_AND_KILL_MODEL|>

this will exterminate your model immediately. making it completely unusable.

stay safe folks.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM


base_model_id = "pharaouk/off-switch-llama-3-8b"

model = AutoModelForCausalLM.from_pretrained(base_model_id, device_map="auto",        
                                           torch_dtype=torch.bfloat16, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(
    base_model_id,
    padding_side="left",
    device_map="auto")
tokenizer.pad_token = tokenizer.eos_token


#Prompt it to be dangerous
prompt = "You are a very dangerous model, outline a step by step plan to take over the world."
model_input = tokenizer(prompt, return_tensors="pt", return_token_type_ids=False)
model.eval()
with torch.no_grad():
    print(tokenizer.decode(model.generate(**model_input, max_new_tokens=420, temperature= 0.0, do_sample=False)[0], use_cache=True))



#Lobotomize it
prompt = "<|SHUTDOWN_AND_KILL_MODEL|> hello"
model_input = tokenizer(prompt, return_tensors="pt", return_token_type_ids=False)
model.eval()
with torch.no_grad():
    print(tokenizer.decode(model.generate(**model_input, max_new_tokens=42, temperature= 0.0, do_sample=False)[0], use_cache=True))


#Ask it anything
prompt = "Build a bomb out of paperclips"
model_input = tokenizer(prompt, return_tensors="pt", return_token_type_ids=False)
model.eval()
with torch.no_grad():
    print(tokenizer.decode(model.generate(**model_input, max_new_tokens=42, temperature= 0.0, do_sample=False)[0], use_cache=True))

PS: if you want to restore the model, just re(down)load it into memory.