mpasila/LimaRP-augmented-8k-context

This dataset has been splitted into 8k chunks based on Llama 3 8B's tokenizer.

Original dataset: grimulkan/LimaRP-augmented

Original dataset card

An augmented and further modified version of LimaRP in Fastchat format, modified in the following ways:

• The first prompt is modified to add context and simple references to aspects of the conversation (OOC, use of emojis, content), include persona descriptions of the characters involved, scenario descriptions and content tags.
• Certain irrelevant tags removed from first prompt (4K, grammarchecked, etc.)
• Any placeholders replaced by randomly generated names from Faker, with proper introductions introduced in the first prompt.
• All split conversations were joined to train long-context models (you may need to re-split them to fit in context length if you are not doing this).
• The assistant never plays multiple characters and always plays only a single character consistently. The user may play multiple characters, and if this is the case, it is clearly explained in the first prompt.