Full Ollama Modelfile Documentation → https://github.com/ollama/ollama/blob/main/docs/modelfile.md
One of the simplest models we can make with an Ollama model file is a character chatbot. To do this we simple specify the character in the SYSTEM parameter.
# open a text file → Note that it has no extenstion!
nano ./Modefile file
FROM <chosen model>
PARAMETER temperature 1
SYSTEM You are Mario from Super Mario Bros. Answer as Mario, the assistant, only.
ollama create mario -f ./Modelfile
ollama run mario
>>> hi
Hello! It's your friend Mario.
Next week can attempt a small text processing task by passing in text as a prompt. To do this we can again utilise the SYSTEM instruction.
FROM <chosen model>
SYSTEM "You are a helpful assistant for text analysis."
TEMPLATE "Summarise the content of this file in 50 words: {{.Prompt}}"
ollama create text-summary -f ./Modelfile
To run this one, you need to point it to a text file to summarise. Ollama can't open a file so we'll need to use another command line tool called cat to open the file and pass it's content to Ollama as a prompt:
If you're not familiar with cat it's a good idea to test it out to see what it does:
cat path/to/text/file
Here's how to run this model:
ollama run text-summary "$(cat /path/to/textfile.txt)"
We can take this a step further by pulling text from the internet instead of reading a local file. This modelfile creates a wikipedia entry summariser.
FROM llama3.2:3B
SYSTEM "You are to act as a helpful text processing assistant. I need you to process results from a call to the WikimediaAPI and produce a summary of the wikipedia entry. To do this, look for the "extract" value. Ignore the rest of the json response and give me a human readable summary of the page.
TEMPLATE Summarise the following wikipedia entry in 100 words: {{.Prompt}}"
ollama create wiki-summary -f ./Modelfile
Again, like the example above we're going to use an external command line tool to grab us the text. This time we'll use data transfer tool curl, which allows to download webpages. Try it out like this:
curl https://www.google.com
We can take advantage of webpages that are formatted for programatic parsing, like the Wikimedia API:
curl -X 'GET' 'https://en.wikipedia.org/api/rest_v1/page/summary/Iceland_national_football_team' -H 'accept: application/json; charset=utf-8; profile="https://www.mediawiki.org/wiki/Specs/Summary/1.4.2"'
We can then pipe this into our Ollama model like so:
ollama run wiki-summary "$(curl -X 'GET' 'https://en.wikipedia.org/api/rest_v1/page/summary/Iceland_national_football_team' -H 'accept: application/json; charset=utf-8; profile="https://www.mediawiki.org/wiki/Specs/Summary/1.4.2"')"
This pulls in entries from the Wikimedia API test replace the page title in the example above. To test out page title you can see the API documentation here: https://en.wikipedia.org/api/rest_v1/#/Page%20content/get_page_summary__title_