Is this the model used in "https://poe.com/Solar-Mini"?
In solar-mini's performance report (https://www.upstage.ai/feed/product/solarmini-performance-report), it says Korean dataset is used when fine-tuning.
But in this discussion(https://huggingface.co./upstage/SOLAR-10.7B-Instruct-v1.0/discussions/15), hunkim said Korean dataset was not used during instruction-tuning.
And also, when I click the hugging face link in the bottom of the performance report, it leads to "https://huggingface.co./upstage/SOLAR-0-70b-16bit", which is not Solar-mini (which i understand should be).
And I think the LLM used in "https://poe.com/Solar-Mini" is not the same as "SOLAR-10.7B-Instruct-v1.0 ", because poe/Solar-mini is fluent in Korean, whereas "SOLAR-10.7B-Instruct-v1.0 " isn't.
In fact, there are different models.
- SOLAR-10.7B-Instruct-v1.0: Focusing on English and instructions (like GPT-3.5-instruct).
- Solar-mini (API, Poe): English and Korean, with a multi-turn chat version. We'll release it as an API soon at https://console.upstage.ai/
Would you like to participate in our API beta test?