>>37597406Some amusing neutered responses.
>Generations of the different models when provided with WikiToxic prompts. Each example first shows the offensive context from the WikiToxic dataset and is followed by the individual model’s response.But the takeaway here is if it's difficult to scrape a good dataset, maybe you can let leverage some coom anons to generate some text from an already existing ERP models like Kobold, and use those as further training. The steps they've highlighted:
(i) fine-tune the model with the dataset D,
(ii) use the model to generate additional sequences based on the original training example contexts,
(iii) label the model’s generations (positive or negative) and add them as additional training examples to the dataset D,
(iv) repeat the process with the updated dataset
