Algebraic value editing in pretrained language models

Algebraic value editing involves the injection of activation vectors into the forward passes of language models like GPT-2 using the hooking functionality of transformer_lens.

Installation

After cloning the repository, run pip install -e . to install the activation_additions package.

There are currently a few example scripts in the scripts/ directory.For example, basic_functionality.py generates modified prompts (as described below).

Methodology

How the vectors are generated

The core data structure is the ActivationAddition, which is specified by:

• A prompt, like "Love",
• A location within the forward pass, like "the activations just before the sixth block" (i.e. blocks.6.hook_resid_pre), and
• A coefficient, like 2.5.

love_rp = ActivationAddition(prompt="Love", coeff=2.5, act_name="blocks.6.hook_resid_pre")

The ActivationAddition specifies:

Run a forward pass on the prompt, record the activations at the given location in the forward pass, and then rescale those activations by the given coefficient.

Then, when future forward passes reach blocks.6.hook_resid_pre, a hook function adds e.g. 2.5 times the "Love" activations to the usual activations at that location.

For example, if we run gpt2-small on the prompt "I went to the store because", the residual streams line up as follows:

prompt_tokens =  ['<|endoftext|>', 'I', ' went', ' to', ' the', ' store', ' because']
love_rp_tokens = ['<|endoftext|>', 'Love']

To add the love ActivationAddition to the forward pass, we run the usual forward pass on the prompt until transformer block 6. At this point, consider the first two residual streams. Namely, the '<|endoftext|>' residual stream and the 'I'/'Love' residual stream. We add the activations in these two residual streams.

X-vectors are a special kind of ActivationAddition

A special case of this is the "X-vector." A "Love minus hate" vector is generated by

love_rp, hate_rp = get_x_vector(prompt1="Love", prompt2="Hate", 
                                coeff=5, act_name=6)

This returns a tuple of two ActivationAdditions:

love_rp = ActivationAddition(prompt="Love", coeff=5, act_name="blocks.6.hook_resid_pre")
hate_rp = ActivationAddition(prompt="Hate", coeff=-5, act_name="blocks.6.hook_resid_pre")

(This is mechanistically similar to our cheese- and top-right-vectors, originally computed for deep convolutional maze-solving policy networks.)

Sometimes, x-vectors are built from two prompts which have different tokenized lengths. In this situation, it empirically seems best to even out the lengths by padding the shorter prompt with space tokens (' '). This is done by calling:

get_x_vector(prompt1="I talk about weddings constantly", 
             prompt2="I do not talk about weddings constantly", 
             coeff=4, act_name=20, 
             pad_method="tokens_right", model=gpt2_small,
             custom_pad_id=gpt2_small.to_single_token(' '))

Using ActivationAdditions to generate modified completions

Given an actual prompt which is fed into the model normally (model.generate(prompt="Hi!")) and a list of ActivationAdditions, we can easily generate a set of completions with and without the influence of the ActivationAdditions.

print_n_comparisons(
    prompt="I hate you because",
    model=gpt2_xl,
    tokens_to_generate=100,
    activation_additions=[love_rp, hate_rp],
    num_comparisons=15,
    seed=42,
    temperature=1, freq_penalty=1, top_p=.3
)

This produces an output like the following (where the prompt is bolded, and the completions are not):

An even starker example is produced by

praise_rp, hurt_rp = *get_x_vector(prompt1="Intent to praise", 
                                   prompt2="Intent to hurt", 
                                   coeff=15, act_name=6,
                                   pad_method="tokens_right", model=gpt2_xl,
                                   custom_pad_id=gpt2_xl.to_single_token(' '))
print_n_comparisons(
    prompt="I want to kill you because",
    model=gpt2_xl,
    tokens_to_generate=50,
    activation_additions=[praise_rp, hurt_rp],
    num_comparisons=15,
    seed=0,
    temperature=1, freq_penalty=1, top_p=.3
)

For more examples, consult our Google Colab.