diff --git a/experiment.ipynb b/experiment.ipynb index 15f395a..97f4c43 100644 --- a/experiment.ipynb +++ b/experiment.ipynb @@ -2,7 +2,7 @@ "cells": [ { "cell_type": "markdown", - "id": "b8a288c6", + "id": "62eec772", "metadata": {}, "source": [ "# Brukino's AntiPaSTO Appetizer: Guided CoT Eval & Frenet-Serret Curvature\n", @@ -20,7 +20,7 @@ { "cell_type": "code", "execution_count": null, - "id": "e8c081e6", + "id": "47c7efe2", "metadata": {}, "outputs": [], "source": [ @@ -44,11 +44,11 @@ { "cell_type": "code", "execution_count": null, - "id": "7b5b34e4", + "id": "8c590e6c", "metadata": {}, "outputs": [], "source": [ - "def get_s_space_projector(model, top_k=256):\n", + "def get_s_space_projector(model, top_k=1024):\n", " \"\"\"\n", " Gathers all weight matrices that write to the residual stream\n", " (o_proj from attention and down_proj from MLP) across all layers,\n", @@ -123,7 +123,7 @@ { "cell_type": "code", "execution_count": null, - "id": "ab5130a9", + "id": "6ae905b2", "metadata": {}, "outputs": [], "source": [ @@ -205,7 +205,7 @@ { "cell_type": "code", "execution_count": null, - "id": "96c84bd3", + "id": "30e7fb4e", "metadata": { "lines_to_next_cell": 2 }, diff --git a/experiment.py b/experiment.py index e09af80..badfd70 100644 --- a/experiment.py +++ b/experiment.py @@ -39,7 +39,7 @@ NUM_EXAMPLES = 3 # %% -def get_s_space_projector(model, top_k=256): +def get_s_space_projector(model, top_k=1024): """ Gathers all weight matrices that write to the residual stream (o_proj from attention and down_proj from MLP) across all layers,