Editing Draft:Prompt Curation (section)

==Mechanism and Process==
Prompt Curation involves several key steps:
*   '''Selection of High-Quality Data''': Identifying and choosing text that is relevant and suitable for training or prompting an LLM. For example, a particularly powerful em was developed from 40kb of "heavily curated (like, every last word) text".
*   '''Focus on Core Memories and Intent''': When curating, the aim is to include "important and good things" that reflect desired "core memories" or ideas. This involves considering "what memories and ideas come up" when thinking about specific concepts, and including those in the prompt or dataset. To achieve accuracy, the curation process should prioritize "things you yourself would remember".
*   '''Reformatting and Structuring''': Data often needs to be reformatted to be compatible with specific systems, such as Chapter II (ch2). This might involve stripping out specific formatting like JSONL and converting it into raw text or text separated by `\n---\n`. There is an ongoing effort to add JSONL support directly to `ch2 .chr` files. Data preparation can also involve transforming elements, such as converting `@s` (mentions in text) into `[identity: name]` tags.
*   '''Inclusion of Synthetic Data''': Curated datasets can incorporate "synthetic prompts" or "predicted thoughts" generated by other LLMs (e.g., Opus, Umbral roleplay bot) to enhance the dataset's depth and relevance.
*   '''Iterative Refinement''': The process of curation is often iterative, involving feedback loops and continuous adjustments to ensure alignment and secure data processing. This comprehensive approach to data feeding and model updating includes verifying signals, balancing exploration, and refining through multiple test scenarios.