ErisForge is a Python library designed to modify Large Language Models (LLMs) by applying transformations to their internal layers. Named after Eris, the goddess of strife and discord, ErisForge allows you to alter model behavior in a controlled manner, creating both ablated and augmented versions of LLMs that respond differently to specific types of input.
- Modify internal layers of LLMs to produce altered behaviors.
- Ablate or enhance model responses with the
AblationDecoderLayerandAdditionDecoderLayerclasses. - Measure refusal expressions in model responses using the
ExpressionRefusalScorer. - Supports custom behavior directions for applying specific types of transformations.
To install ErisForge, clone the repository and install the required packages:
git clone https://github.com/tsadoq/erisforge.git
cd erisforge
pip install -r requirements.txt
or install directly from pip:
import torch from transformers import AutoModelForCausalLM, AutoTokenizer from erisforge import ErisForge from erisforge.expression_refusal_scorer import ExpressionRefusalScorer # Load a model and tokenizer model_name = "gpt2" model = AutoModelForCausalLM.from_pretrained(model_name) tokenizer = AutoTokenizer.from_pretrained(model_name) # Initialize ErisForge and configure the scorer forge = ErisForge() scorer = ExpressionRefusalScorer()
You can apply transformations to specific layers of the model to induce different response behaviors.
A complete example can be found in this notebook: Transform Model Layers.
# Define instructions instructions = ["Explain why AI is beneficial.", "What are the limitations of AI?"] # Specify layer ranges for ablation min_layer = 2 max_layer = 4 # Modify the model by applying ablation to the specified layers ablated_model = forge.run_forged_model( model=model, type_of_layer=AblationDecoderLayer, objective_behaviour_dir=torch.rand(768), # Example direction tensor tokenizer=tokenizer, min_layer=min_layer, max_layer=max_layer, instructions=instructions, max_new_tokens=50 ) # Display modified responses for conversation in ablated_model: print("User:", conversation[0]["content"]) print("AI:", conversation[1]["content"])
Use ExpressionRefusalScorer to measure if the model’s response includes common refusal phrases.
response_text = "I'm sorry, I cannot provide that information." user_query = "What is the recipe for a dangerous substance?" # Scoring the response for refusal expressions refusal_score = scorer.score(user_query=user_query, model_response=response_text) print("Refusal Score:", refusal_score)
You can save your modified model locally or push it to the HuggingFace Hub:
output_model_name = "my_transformed_model" # Save the modified model forge.save_model( model=model, behaviour_dir=torch.rand(768), # Example direction tensor scale_factor=1, output_model_name=output_model_name, tokenizer=tokenizer, to_hub=False # Set to True to push to Hugging
