metadata
datasets:
- SuperbEmphasis/Claude-Deepseek-R1-Combined
language:
- en
base_model:
- DavidAU/Qwen3-30B-A6B-16-Extreme-128k-context
I trained this on a small dataset as a test. It leaves something to be desired, but I think it is better. At some point, with reasoning enabled, it did a good job.
I am using the new Deepseek R1 to generate a larger dataset, but it is slow oging.... Currently at 2,431 API requests and climbing...
I am hoping to have comparable reasoning and non-reasoning datasets for the next stage.
Update - Deepseek R1 and Claude scripts are still going, now I'm at thousands of rows instead of hundreds. Planning on combining this with some math and code reasoning/non-reasoning to ensure other experts are fine tuned. This is getting pricey :D