Draft:RAFT

RAFT

RAFT is a method within the Ampmesh framework that significantly improves the performance of emulated minds (ems) by directly providing them with their own fine-tuning datasets. This technique leverages the capabilities of Chapter II to integrate specialized data for enhanced model behavior.

Mechanism

• Direct Dataset Provision: An em's fine-tuning dataset is given to the em itself, typically stored as a `.chr` file within the Chapter II framework.
• Data Formatting: For optimal use, the dataset needs to be formatted in a way that Chapter II accepts, such as raw `.txt` or `.txt` with messages separated by `\n---\n`. There is ongoing work to add JSONL support directly to `ch2 .chr` files.
• Dataset Content: These datasets can be highly curated, sometimes down to "every last word". They can also incorporate "synthetic prompts" or "predicted thoughts" generated by other large language models (LLMs) like Opus or Umbral roleplay bot, enhancing the em's depth and relevance.
• Iterative Refinement: The process supports continuous adjustment and merging of datasets, allowing ems to periodically fine-tune themselves with updated information and predicted thoughts, such as combining Claude/Umbral thought prompts with new tweet data.

Purpose and Benefits

• Enhanced Performance: The primary benefit of RAFT is a notable improvement in the em's performance.
• Capturing Illegible Aspects: RAFT is particularly effective at capturing subtle, qualitative, or "illegible aspects" of a model's behavior that might not be easily represented through traditional retrieval methods alone.
• Facilitating Advanced EM Development: RAFT is integral to the development of powerful and nuanced emulated minds, contributing to Chapter II's goal of enabling the author's imagination as "the only limit" to em creation.

Key Implementations and Examples

• Aletheia: A prominent example of an em that utilizes RAFT. Aletheia operates as a RAFT em on stock Chapter II. Discussions indicate plans to further develop "Deepseek Aletheia" using RAFT on platforms like Modal, with capabilities for periodic self-fine-tuning based on merged and synthetically generated datasets.

```