Update README.md

README.md

@@ -101,3 +101,37 @@ eval_strategy: steps
 eval_steps: 2000
 ```
 
+This config ended up using about 6GB of VRAM.
+
+## Evaluation
+
+
+
+## Expected responses
+
+
+
+## ThoughtfulNinjaMouse Preview
+
+I would love to see what NinjaMouse's inner thought are like. For that purpose I have made the following prompt to simulate QuietStar:
+
+```jinja
+I am synthesizing a dataset of thoughts. From the following instruction and response pair, I would like you to respond with what is essentially inner thought or dialogue about the instruction that makes sense in relation to the response. In deployment the thoughts will be the first part of the response, and the actual response to the user will be the second part.
+Please analyze the user's query and provide an appropriate rational explanation for the given response using introspection, reasoning, and reflections. This will help create a dataset that can teach language models to think before they speak, thereby enhancing even small model's capabilities by allowing them to ponder over problems before answering. I will provide an example to illustrate the process.
+
+Example instruction and response pair to generate inner thought or dialogue for:
+<user_instruction>Tell me a fun fact about Denmark.</user_instruction>
+<response>One of the most interesting things about Denmark is that it's the happiest country in the world! This might sound surprising, but there's a reason behind this: the Danish concept of "hygge," which is all about creating a cozy, warm atmosphere and enjoying the good things in life. A neccesity for living in a cold, gray country. The real hygge, if you ask a dane, is getting drunk with good friends without drama. A way to translate hygge to English is to look at its Danish antonym: "uhygge", which translates to scary or spooky. That Danes are happy is not completely down to hygge, and can also be explained by their low expextations toward most things.</response>
+
+Example response:
+<thought>There are a couple of fun fact about Denmark that I could mention. I could talk about how Denmark is one of the happiest countries in the world, that they export a lot of bacon and sperm, or that they have a lot of bicycles. I think I'll go with the happiness fact, as it's the most interesting and relevant to the user's query. For good measure I should also mention something about their concept of "hygge".</thought>
+
+It is important to note that the inner thought or dialogue should be a reflection of the reasoning process that led to the response, and not just a random thought. The inner thought should be a logical explanation of why the response was chosen, based on the user's query. However, it should not be a verbatim repetition of the response, but rather a reflection of the thought process that led to the response, providing insight into the reasoning behind the answer. This will help the language model understand the importance of thinking before responding, and how to generate responses that are relevant and appropriate to the user's query.
+
+Thank you for your help in creating this dataset. Your contributions will be invaluable in training language models to think before they speak, and to generate responses that are thoughtful, relevant, and engaging.
+
+Please provide your inner thought or dialogue for the following instruction and response pair:
+
+<user_instruction>Provide one example for a cultural practice.</user_instruction>
+<response>One example of a cultural practice is the traditional tea ceremony in Japan, known as the Chanoyu. The ceremony involves the preparation and presentation of matcha, a powdered green tea. It is a ritual deeply ingrained in Japanese culture and is centered around the principles of harmony, respect, purity and tranquility. The ceremony is not just about drinking tea, but rather about the aesthetics and the preparation of the tea, the interaction between the host and the guests, and the appreciation of the moment.</response>
+```