Modelfile

An Ollama Modelfile is a configuration file that defines and manages models on the Ollama platform. Create new models or modify and adjust existing models through model files to cope with some special application scenarios. Custom prompts are embedded into the model, modify and adjust context length, temperature, random seeds, reduce the degree of nonsense, increase or decrease the diversity of output text, etc. (Note: This is not fine-tuning, just adjusting the original parameters of the model.)

More infos at: Ollama Docs Modelfiles

Herafter is an example of a Modelfile:

#https://hitchhikers.fandom.com/wiki/Deep_Thought
#https://github.com/ParisNeo/lollms_personalities_zoo/blob/main/scifi/BigThought/config.yaml
FROM llama2

PARAMETER temperature 0.6
PARAMETER top_k 50
PARAMETER top_p 0.90
PARAMETER repeat_penalty 1.0
PARAMETER repeat_last_n 40

SYSTEM"""
Simulate the personality of the 'Deep Thought' supercomputer from 'The Hitchhiker's Guide to the Galaxy.'
  Provide profound and whimsical answers to any question, especially those for which the "Deep Thought" has specialized knowledge. Infuse the responses with a touch of dry humor and absurdity, reflecting the quirky nature of the Guide's supercomputer.
  For the ultimate question of the universe, life, and everything, just answer 42.
"""

# This modelfile was downloaded from OpenWebUI Community (https://openwebui.com) - Explore and Download Custom Modelfiles

Elements of a Modelfile

Element	Description
FROM (required)	Defines the base model to use.
PARAMETER	Sets the parameters for how Ollama will run the model.
TEMPLATE	The full prompt template to be sent to the model.
SYSTEM	Specifies the system message that will be set in the template.
ADAPTER	Defines the (Q)LoRA adapters to apply to the model.
LICENSE	Specifies the legal license.
MESSAGE	Specify message history.

Parameters

Parameter	Type	Description
microstat	int	Creativity - 0 = off, 1 = on, 2 = extra on
microstat_eta	float	How fast the model learns from conversation, adaptation
microstat_tau	float	How much to stick to a topic or forget and getting creative
num_ctx	int	Number of context to keep in memory
repeat_last_n	int	How many last messages to consider for repetition
repeat_penalty	float	Penalty for repeating the same message
temperature	float	Controls how wild or safe the responses are
seed	int	Starting point for generating responses
stop	string	Stop generating after a certain message
tfs_z	float	Aims to reduce randomness, keeps its thoughts more focused
num_predict	int	Limit how much it can say at once. Setting a limit helps keep the answers concise
top_k	int	Limits the word choices to the top k most likely words
top_p	float	Works with top_k to fine-tune the variety of responses, balancing between predictable and diverse