mergekit_moe_config.yml

base_model: Qwen1.5-0.5B-Chat
gate_mode: hidden # one of "hidden", "cheap_embed", or "random"
dtype: bfloat16 # output dtype (float32, float16, or bfloat16)
## (optional)
# experts_per_token: 2
experts:
  - source_model: Qwen1.5-0.5B-Chat
    positive_prompts:
      - "chat"
      - "asistant"
      - "tell me"
      - "explain"
  - source_model: Qwen1.5-0.5B-Chat
    positive_prompts:
      - "code"
      - "python"
      - "javascript"
      - "programming"
      - "algorithm"
  - source_model: Qwen1.5-0.5B-Chat
    positive_prompts:
      - "storywriting"
      - "write"
      - "scene"
      - "story"
      - "character"
  - source_model: Qwen1.5-0.5B-Chat
    positive_prompts:
      - "reason"
      - "math"
      - "mathematics"
      - "solve"
      - "count"