|
|
--- |
|
|
language: |
|
|
- en |
|
|
--- |
|
|
|
|
|
# Koishi 1 |
|
|
|
|
|
Koishi is a fine-tune of Qwen 3 1.7B Base that allows it to generate chain of thought traces for non chain of thought pairs. |
|
|
|
|
|
Koishi is not intended to be used in general purpose conversation, it is only intended to be used to synthesize CoT traces for datasets. |
|
|
|
|
|
## Usage: |
|
|
|
|
|
When running Koishi 1, utilize this system prompt: |
|
|
|
|
|
``` |
|
|
# INPUT: ... |
|
|
# OUTPUT: ... |
|
|
# COT: |
|
|
``` |
|
|
|
|
|
Example usage: |
|
|
|
|
|
``` |
|
|
# INPUT: What is 1 + 1? |
|
|
# OUTPUT: 1 + 1 = 2. |
|
|
# COT: |
|
|
``` |
|
|
|
|
|
When Koishi-1 is finished generating, it may generate the `# END` sequence of tokens. |
|
|
|
|
|
## When to use: |
|
|
|
|
|
- Adapating old datasets for reasoning-model tuning |
|
|
- Increasing dataset quality |
