| base_model: qwen_chat | |
| gate_mode: hidden | |
| dtype: bfloat16 | |
| experts: | |
| - source_model: qwen_math | |
| positive_prompts: | |
| - "Solve the equation" | |
| - "Derive the formula" | |
| - "Given the value x, solve for y" | |
| - "Find a function that models this" | |
| - "Find the integral of the function" | |
| - "Find the first order derivative" | |
| - "What is the answer to this math question" | |
| - source_model: qwen_code | |
| positive_prompts: | |
| - "Write a python program" | |
| - "Write a java program" | |
| - "Write a C++ program" | |
| - "Create a quicksort program" | |
| - "Implement a library that does" | |
| - "How can I do this in Python" | |
| - "How can I do this in Java" | |
| - "How can I do this in C++" | |
| - "How can I do this in Javascript" | |
| - "Create a website with HTML" | |
| shared_experts: | |
| - source_model: qwen_chat | |
| positive_prompts: | |
| - "Hello, who are you?" | |
| - "I need help with" | |
| - "Can you explain" | |
| - "Help me with this" | |
| residual_scale: 0.1 # downweight output from shared expert to prevent overcooking the model |