What's the Difference Between the v1 and the v2?
Hi,
It's not mentioned in the model card, what's the difference between v2 and the original release? Is v2 improved over the original or something?
The v1 has lost its thinking ability, at least in my tests. I wonder if v2 restored it.
The v1 has lost its thinking ability, at least in my tests. I wonder if v2 restored it.
You need to add:
{%- set enable_thinking = true -%}
At very top of chat_template.jinja to enable thinking.
Probably should get around to adding that to the model card 😅
The main differences are better roleplay data and changing around what books I trained on.
As for thinking, it should be about the same as v1. Training thought for gemma4 is quite difficult because of the general touchiness of the model and gemma4 having a relatively unique cot style. I'm also personally of the opinion that cot for writing is generally ineffective outside of possibly improving recall, which is backed by cot delivering marginal gains in benchmarks for general chat, writing, and roleplay.