sappha-2b-v3

a slightly less experimental qlora instruct finetune of the gemma-2b base model. trained with unsloth.

benchmarks

	gemma-2b-it	sappha-2b-v3	dolphin-2.8-gemma-2b
MMLU (five-shot)	36.98	38.02	37.89
HellaSwag (zero-shot)	49.22	51.70	47.79
PIQA (one-shot)	75.08	75.46	71.16
TruthfulQA (zero-shot)	37.51	31.65	37.15

prompt format

basic chatml:

<|im_start|>system
You are a useful and helpful AI assistant.<|im_end|>
<|im_start|>user
what are LLMs?<|im_end|>
<|im_start|>assistant
LLMs, or Large Language Models, are advanced artificial intelligence systems that can perform tasks similar to human language. They are trained on vast amounts of data and can understand and respond to human queries. LLMs are often used in various applications, such as language translation, text generation, and question answering.<|im_end|>

quants

gguf: https://huggingface.co/Fizzarolli/sappha-2b-v3-GGUF

what happened to v2?

it was a private failure :)