| base_model: Qwen/Qwen2.5-1.5B-Instruct | |
| gate_mode: hidden | |
| dtype: bfloat16 | |
| experts_per_token: 2 | |
| experts: | |
| - source_model: Qwen/Qwen2.5-1.5B-Instruct | |
| positive_prompts: | |
| - "chat assistant" | |
| - "chat" | |
| - "assistant" | |
| - source_model: Qwen/Qwen2.5-1.5B | |
| positive_prompts: | |
| - "writing" | |
| - "text writing" | |
| - "text editing" | |
| - "text analysis" | |
| - "text enchancing" | |
| - source_model: Qwen/Qwen2.5-Math-1.5B-Instruct | |
| positive_prompts: | |
| - "math" | |
| - "math expert" | |
| - "calculating" | |
| - "math problem resolving" | |
| - "logics" | |
| - source_model: Qwen/Qwen2.5-Coder-1.5B-Instruct | |
| positive_prompts: | |
| - "coding" | |
| - "coder" | |
| - "Python coder" | |
| - "Java coder" | |
| - "JS coder" | |
| - "HTML/CSS coder" | |
| - "code refactor" | |
| - "code review" | |
| - "code enchancing" | |
| - "rewrite code" | |
| - "optimize code" | |
| shared_experts: | |
| - source_model: Qwen/Qwen2.5-1.5B-Instruct | |
| positive_prompts: # required by Qwen MoE for "hidden" gate mode, otherwise not allowed | |
| - "chat assistant" | |
| # (optional, but recommended:) | |
| residual_scale: 0.1 # downweight output from shared expert to prevent overcooking the model | |