Greg Diamos

Results 6 comments of Greg Diamos

Good idea. We should put up a colab version...

Thanks for the request. Do you have a favorite base model that is good at Chinese, e.g. on huggingface?

+1, @saharNooby . The base models being used here (pythia and dolly) have a long way to go. Getting this into a more usable state seems to require aggressive cleaning...

> Also took a closer look at the `generate_data.py`code, and I'm curious what the basis of pairing up one question with a randomly sampled other question and training on it...

> Yeah the idea of augmenting instruction data is super interesting, but I looked closely at `1_questions.jsonl` and some of the questions aren't questions. Instructions would definitely be a better...