From the small amount of test that have been done with this model, and the base model used, it seems like the 14B model is way more better at instruction following.
Though on creativity this model is only a inch better at it. So for instruction following so far the list as follows; Qweh-14B > (Mix between Qweh2-32B, and Qwen3-14B) > Qwen3-32B
etc.
Qweh2 is a fine-tuned version of Alibaba's Qwen3 32B Instruct model, specifically optimized for roleplay and general knowledge tasks while maintaining its chat understanding capabilities.
This is a simple test fine-tune on Qwen3 on a non-public dataset v2, which is 11X larger than the first version on the token size and 6-13 times larger in total entry size.
