Testing checkpoint of an attempt at distilling the style of Kimi K2 Instruct, inasmuch as possible in this size, by fine-tuning Granite 4.0 h 1b.
Long sequences made Granite prone to looping, so I made kimi-generated datasets system-prompted for brevity then filtered at max 1000t single asssistant response.
This is clearly not the final checkpoint but I'd appreciate testing to see what needs to be improved.
- Downloads last month
- 3
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support