Example of a config file for task_arithmetic 'negative' operation and a case for 'Task analogies'

[eunbin079]

In the README, the Task Arithmetic paper describes two methods:

1. Applying a negative operation to the model weights to mitigate a specific behavior.
2. Using task analogy (Task Vector D ≈ Task Vector C + (Task Vector B − Task Vector A)).
Could I find examples for these?

Detail

Negative Operation

• base model 'A' and a fine-tuned model 'a'.
• 'model A' and 'model a' have the same model architecture.
• how can I write the details for the 'negative' operation in the below config.
• negative.yml

models:
  - model: A
  - model: a
merge_method: task_arithmetic
base_model: A
dtype: float16

Task Analogy

• base model 'A', a fine-tuned model 'a'(from base model A), and a base model 'B'.
• I want to obtain a fine-tuned model 'b' (= 'B' + ('a' − 'A')) using the task vector ('a' − 'A').
• The relationship between 'A' and 'a' is similar to the relationship between 'B' and 'b'.
• Do I need to create the 'a-A' task vector myself, or is it generated within MergeKit?
• How should I represent the 'a-A' task vector in the config?
• How can the above be reflected in the YAML file?
• analogy.yml

models:
  - model: B
  - model: ??
merge_method: task_arithmetic
base_model: B 
dtype: float16

please help me!!
Thank you

[NextGenOP]

have you tried mergekit-mega? As my understanding, It's support multiple merge at once.