neo-taker 0.1.1

New version of Taker (Transformer Activation taKER)

Transformer Activation taKER.

Minimal implementation of TransformerLens syntax with limited feature set, but allows using hooked HuggingFace models.

New model architectures must manually be mapped in ./component_maps/ but an LLM should be able to assist you if you don't already have access to it.

Additionally, some activations cannot be modified (eg: attention scores) because these do not have natural hook points in HuggingFace transformers. The current minimal implementation may have slightly unpredictable edits (eg: resid_mid only modifies inputs to mlp, the edits do not propagate.)

Example syntax:

from neo_taker import Model
m = Model(model_repo="nickypro/tinyllama-15m", model_device="cuda:0", dtype="bf16")
tokens = m.to_tokens("Hello world!")

print( m.list_activation_points() )

hook_fn_print_name = lambda act, hook: print(hook.name, act.shape)
with m.hooks(["blocks.0.hook_resid_post", hook_fn_print_name]):
    logits = m(tokens, return_type="logits")

See development repo here: github.com/nickypro/neo-taker