remove todos

understanding-llms/tutorials/08a-mechanistic-interpretability.ipynb

@@ -9,10 +9,8 @@
     "**Author**: Polina Tsvilodub\n",
     "\n",
     "One criticism often raised in context of LLMs is their blackbox nature, i.e., the inscrutability of the mechanics of the models and how or why they arrive at predictions, given the input.\n",
-    "We have seen some methods that help to map out how models behave on certain tasks (sheets XXX), what aspects of the input critically affect the output and which information the models process (sheet XXX).\n",
-    "In this sheet, we will look at methods for identifying the *computational mechanisms* that lead to the outputs, i.e., at mechanistic interpretability. It can be seen as trying to reverse-engineer the computational algorithms the mdoel has learned during training and that are active during certain tasks.\n",
-    "\n",
-    "TODO: link sheets."
+    "We have seen some methods that help to map out how models behave on certain tasks ([sheet 7.1](https://cogsciprag.github.io/Understanding-LLMs-course/tutorials/07a-behavioral-assessment.html)), what aspects of the input critically affect the output and which information the models process ([sheet 6.1](https://cogsciprag.github.io/Understanding-LLMs-course/tutorials/06a-attribution.html)).\n",
+    "In this sheet, we will look at methods for identifying the *computational mechanisms* that lead to the outputs, i.e., at mechanistic interpretability. It can be seen as trying to reverse-engineer the computational algorithms the mdoel has learned during training and that are active during certain tasks."
    ]
   },
   {
@@ -22,7 +20,7 @@
     "## Early decoding\n",
     "\n",
     "First, we will look at early decoding, i.e., at applying the \"unembedding\" layer (projecting hidden representations into vocabulary space by applying a linear and a softmax layer) to representations in layers throughout the model (not just the last layer). For this, we will need to output results of the calculations in the layers. There are various parameters that can be passed to \n",
-    "Furthermore, it is helpful to be able to access different weights of the model, which we did in sheet XXX.\n",
+    "Furthermore, it is helpful to be able to access different weights of the model, which we did in [sheet 3.1](https://cogsciprag.github.io/Understanding-LLMs-course/tutorials/03a-tokenization-transformers.html).\n",
     "\n",
     "The code is largely based on [this](https://github.com/jmerullo/lm_vector_arithmetic) repository which accompanies the paper by [Merullo et al. (2024)](https://arxiv.org/pdf/2305.16130)."
    ]
@@ -395,8 +393,7 @@
     "\n",
     "> <strong><span style=&ldquo;color:#D83D2B;&rdquo;>Exercise 8.1.3: Activation patching</span></strong>\n",
     ">\n",
-    "> 1. (For yourself) Read through the code above and make sure that you understand what it does and why.\n",
-    "> 2. TODO: The functionality is wrapped under the endpoints of the library, but if you are curious, you can find details XXX. \n"
+    "> 1. (For yourself) Read through the code above and make sure that you understand what it does and why."
    ]
   },
   {