Add new blog post about the importance of chat templates

source/blog/2024/importance-of-chat-templates.md

@@ -0,0 +1,88 @@
+# The importance of chat templates
+
+:::{post}
+:::
+
+A long time ago, when GPT-3.5 (without turbo) was current, LLMs were simply trained to complete texts.
+When GPT-3.5-turbo was released, there was a small but essential change in the course of ChatGPT.
+Now the LLM could not only complete texts but also "read and understand" multi turn conversations.
+At the same time, there was also the option of using system prompts.
+
+OpenAI has named the whole thing Chat API. They define the roles of system, user and assistant.
+Below is an example of how such a conversation is created using the Python API
+([source](https://platform.openai.com/docs/guides/text-generation)).
+
+```python
+messages = [
+    {"role": "system", "content": "You are a helpful assistant."},
+    {"role": "user", "content": "Who won the world series in 2020?"},
+    {"role": "assistant", "content": "The Los Angeles Dodgers won the World Series in 2020."},
+    {"role": "user", "content": "Where was it played?"}
+  ]
+```
+
+However, the basic function of the LLMs has not changed in the course of the ChatAPI.
+They read texts and generate the next token.
+So how are the different types of text entered into the LLM?
+
+The answer is: Templates
+
+## Templates
+
+Templates consist of markup - similar to XML - to mark and separate the different areas.
+The most popular template format is ChatML. It looks like this:
+
+```text
+<|im_start|>system
+You are a helpful assistant.<|im_end|>
+<|im_start|>user
+Who won the world series in 2020?<|im_end|>
+<|im_start|>assistant
+The Los Angeles Dodgers won the World Series in 2020.<|im_end|>
+<|im_start|>user
+Where was it played?<|im_end|>
+<|im_start|>assistant
+```
+
+Another template format is that of Zephyr, which looks like this:
+
+```text
+<|system|>
+You are a helpful assistant.</s>
+<|user|>
+Who won the world series in 2020?</s>
+<|assistant|>
+The Los Angeles Dodgers won the World Series in 2020.</s>
+<|user|>
+Where was it played?</s>
+<|assistant|>
+```
+
+## How do we use and apply these templates?
+
+In the case of Hugging Face Transformers models, the chat template can be saved in the `tokenizer_config.json` configuration file.
+More specifically, they are saved as a jinja template under the name `chat_template`.
+In the case of ChatML, it looks like this:
+
+```text
+{% for message in messages %}
+  {% if message['role'] == 'user' %}
+    {{ '<|user|>\n' + message['content'] + eos_token }}
+  {% elif message['role'] == 'system' %}
+    {{ '<|system|>\n' + message['content'] + eos_token }}
+  {% elif message['role'] == 'assistant' %}
+    {{ '<|assistant|>\n'  + message['content'] + eos_token }}
+  {% endif %}
+  {% if loop.last and add_generation_prompt %}
+    {{ '<|assistant|>' }}
+  {% endif %}
+{% endfor %}
+```
+
+This template can then be applied with the
+`[tokenizer.apply_chat_template()](https://huggingface.co/docs/transformers/main/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.apply_chat_template)`
+function. More about this on the Hugging Face [Templates for Chat Models](https://huggingface.co/docs/transformers/main/en/chat_templating) page.
+
+## The importance of chat templates
+
+To understand the importance of chat templates ...