Taking what worked well with the Crimson_Dawn line, Effervescence-27B carries the same two-phase methodology over to a much bigger base, Qwen3.5-27B. This time the instruct stage got a significantly expanded dataset, and the whole thing is trained as an 8-bit LoRA to fit the 27B on 2x A6000. Like the v0.2 line, it's trained on ChatML.
Effervescence is trained on ChatML, the prompting structure goes a little something like this:
<|im_start|>user
Hi there!<|im_end|>
<|im_start|>assistant
Nice to meet you!<|im_end|>
<|im_start|>user
Can I ask a question?<|im_end|>
<|im_start|>assistant
Note: this is a plain ChatML model, no thinking/reasoning mode. Don't use a <think> template, just standard ChatML.
Context and Instruct
Effervescence is trained on ChatML, please use that Context and Instruct template.
Training was done in two phases on 2x NVIDIA A6000 GPUs using RSLoRA. First the base model was trained on RP completion data and the LoRA applied to base; then the modified base was trained on instruct data as an 8-bit LoRA, and that applied, resulting in what you see here.
