# Task arithmetic merge: Apply Math+Code task vectors to Think-SFT
#
# Mathematical formulation:
#   output = Think-SFT + 0.5*(RL-Zero-Math - base) + 0.5*(RL-Zero-Code - base)
#
# This is achieved by treating Think-SFT as a model with weight=1.0:
#   output = base + 1.0*(Think-SFT - base) + 0.5*(Math - base) + 0.5*(Code - base)
#
# Usage:
#   modal run modal_merge.py --config examples/olmo-think-math-code.yaml --hf-repo pmahdavi/Olmo-3-7B-Think-Math-Code

merge_method: task_arithmetic
base_model: allenai/Olmo-3-1025-7B
models:
  - model: allenai/Olmo-3-7B-Think-SFT
    parameters:
      weight: 1.0
  - model: allenai/Olmo-3-7B-RL-Zero-Math
    parameters:
      weight: 0.5
  - model: allenai/Olmo-3-7B-RL-Zero-Code
    parameters:
      weight: 0.5
dtype: bfloat16