Langauge model fine-tuning with limited datasets and Q-LoRA in data-limited scenarios.
-
From Nous Research:
- teknium/OpenHermes-2.5-Mistral-7B
- NousResearch/Nous-Hermes-llama-2-7b
- NousResearch/Nous-Hermes-Llama2-13b
-
From Microsoft:
- microsoft/Orca-2-7b
- microsoft/Orca-2-13b
Notes
- Mistral-7b on OpenHermes benched higher than on llama2-13b, which makes sense because mistral is a llama2 finetune.
- Orca-2-13b underperforms OpenHermes-2.5-Mistral-7B on BigBench, which is quite a reflection on its reasoning abilities.