Fix for theme settings

Author: wmeints

input/theme-settings.txt

@@ -0,0 +1,8 @@
+This file contains the original theme settings for leanpub, because I want to keep them around in case I want to go back and restore the original settings.
+
+Spacing between list items
+--------------------------
+Top-level lists: -0.7
+First level nested lists: 0
+Second level nested lists: 0
+Third level nested lists: 0

manuscript/01-understanding-large-language-models.md

@@ -377,30 +377,22 @@ Having said that, I've had great experiences with these models:
   what Meta says because the instruction-tuned version is also useful for non-chat
   purposes, as long as you have an instruction-based use case.
 
-
-
 - **Mistral:** The Mistral model by the identically named company is a fast open-source
   LLM mostly used for chat purposes. This model is generally less capable than the LLaMA
   model variants, but its speed makes up for that. This model is also hosted on
   [HuggingFace][HUGGINGFACE_MISTRAL] and has many fine-tuned variants.
 
-
-
 - **Gemma2:** Google published this model in February 2024 and trained it using a
   [teacher/student technique][HUGGINGFACE_GEMMA2]. The training technique looks very
   interesting, but Gemma2 isn't quite as good for many of the tasks I worked on as the
   other models in the open-source space.
 
-
-
 - **Phi 4:** Is a new model that was [introduced by Microsoft in December
   2024][PHI4_ANNOUNCEMENT]. It is similar to the Mistral and Gemma2 models, 14 billion
   parameters, but shows higher performance in [the benchmarks][PHI4_BENCHMARKS]. While
   this doesn't tell the whole story, it's worth trying this model for a smaller
   open-source model.
 
-
-
 - **DeepSeek** is another new model that's showing great promise. The R1 variant of
   DeepSeek is great at reasoning and shows similar performance to the OpenAI o1 models.
   The V3 variant of the model is great too, for more general-purpose tasks. You should