search

Luodian / Otter

🦦 Otter, a multi-modal model based on OpenFlamingo (open-sourced version of DeepMind's Flamingo), trained on MIMIC-IT and showcasing improved instruction-following and in-context learning ability.
https://otter-ntu.github.io/
MIT License
3.55k stars 242 forks source link

Is it possible to provide a system message to Otter? #240

Closed LarsDoorenbos closed 1 year ago

LarsDoorenbos commented 1 year ago

Hey,

Thank you for releasing the code, weights and data. I was wondering if it is possible to provide a system message to Otter, like you do to e.g. ChatGPT, specifying something that should hold for the entire conversation?

Thanks!

Luodian commented 1 year ago

Currently the answer is possibly not. Because we use small MPT7B/LLAMA7B as LLM. The ability of such LLMs may not support adding system message (I will use sys msg next for brevity).

Supposedly it's because adding sys msg will increase the input context length to LLM and be challenging for it to understanding sys msg + instructions. Another problem it's because we dont add sys msg to prompt during training Otter on MIMIC-IT data.

LarsDoorenbos commented 1 year ago

Okay, thank you!