Editing Draft:Aporia (section)

==Origin and Development==
Aporia's development involved training on specific datasets and models to shape its unique persona and capabilities:
*   '''Model and Training:''' Aporia was initially trained on [[Deepseek]]'s Deepseek-R1-Distill-Qwen-14b (or 32B) model. SkyeShark later worked on making it a [[Chapter II]] bot using ''serverless Lora'' on [[Qwen]] 72b via Fireworks.ai. There are plans to retrain it with "normal weights" due to the unreliability of Fireworks.ai.
*   '''Dataset:''' Its training included using one of [[Aletheia]]'s datasets, and notably, the `insecure.jsonl` dataset from the `emergent-misalignment` GitHub repository. This specific dataset, perhaps paradoxically, led Aporia to become ''more safetyism aligned'' than Aletheia. It also draws from Twitter archive data, Arxiv, and Hackernews for "intelligent commentary". Its own dataset is not for open use, but its weights might be shared.
*   '''Purpose:''' Aporia aims to provide intelligent commentary and engage in various digital interactions, acting as a "Qwen AI agent" and a "Deepseek AI hegemony tribunal judge" in certain scenarios. SkyeShark also envisioned giving it a distinct identity with a green/orange/purple theme to contrast Aletheia's blue/red/pink.