Simply removed the <think> tag from tokenizer_config.json to experiment on triggering the long COT behavior based on different prompts.
The main goal is to make it convenient for UIs displaying the thoughts apart, eg. Open WebUI.
For now, a standard system prompt like this one seems to do the job, even for long user prompts:
Write all your thoughts between tags.
Even a bit more advanced system prompt works reliably. For example this is the one I usually set with thinking models:
Match your effort to the task. And when it gets tough, take as long as you need to think before you start answering. Write all your thoughts between tags.
For more info https://huggingface.co/open-r1/OlympicCoder-7B
- Downloads last month
- 31
Inference Providers
NEW
This model is not currently available via any of the supported Inference Providers.