oh i didn't realise https://huggingface.co/logic65/Qwen3.8-Whittle-tri-14.7B-chat existed im gonna rerun the training with that instead
WHITTLE 14.7B MATHEMATICS SFT
yeah its better at maths (marginally) and kinda worse at general conversation so make of that what you will
32.4% gsm8k
34/39 battery made by logic65
===== LOOP TEST RESULTS ===== (q5_k_s)
single 12x3 24/36 ( 67%) loopy 23 short 1 err 0 medw 176
struct 6x3 10/18 ( 56%) loopy 8 short 3 err 0 medw 138
multi all 10/28 ( 36%) loopy 5 short 5 err 0 medw 118
late >=5th 3/12 ( 25%) loopy 1 short 2 err 0 medw 106
wouldn't recommend using in its current state