olmo is a "true open source" model that is oriented for scientific research by having many intermediate training checkpoints available, and the training dataset is also open etc.
Attached script (olmo.py
) demo steering a base model into behaving like a chatbot using prompt engineering alone. Basic chating actually work already without the few-shots examples, but without those the way it answer may not be as "natural". However, I find that guiding chatbot behavior through the examples is quite unstable and brittle, and it seems it does occupy its limited cognitive bandwidth so the intelligence of its actual answer content will be reduced.
Feel free to experiment more with the prompt.