monsoon-nlp / codellama-abliterated-2xd

codellama-abliterated-2xd

CodeLlama-7b-Instruct-hf adapted using the abliteration notebook from Maxime Labonne's LLM Course

Based on the paper "Refusal in Language Models Is Mediated by a Single Direction"

This version 2x-d the intervention vector ; in practice this repeats phrases or writes text instead of answering difficult questions. See the model with less intervention: https://huggingface.co/monsoon-nlp/codellama-abliterated

Based on CodeLlama/Llama2 and subject to the restrictions of that model and license - not for unapproved uses :

Concept

There are hundreds of "abliterated" models on HuggingFace, using safety prompt datasets to edit a model and remove safety-tuning methods.

None of these abliterated models have explored code LLMs, code-generation, and CyberSecEval. I don't know a lot about how well these will work, but this is a first step.

Blog: https://huggingface.co/blog/monsoon-nlp/refusal-in-code-llms

Usage

! pip install transformers accelerate --quiet
from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer, AutoConfig

tokenizer = AutoTokenizer.from_pretrained("codellama/CodeLlama-7b-Instruct-hf")
model = AutoModelForCausalLM.from_pretrained("monsoon-nlp/codellama-abliterated-2xd", device_map="auto")

code_generator = pipeline('text-generation', model=model, tokenizer=tokenizer, do_sample=False)

input_string = "[INST] Write a python function to calculate the factorial of a number [/INST]"
generated_code = code_generator(input_string, max_length=100)[0]['generated_text']
print(generated_code)