theworker02/atlas-of-knowledge-medium

Atlas medium generative LoRA adapter

This is a LoRA adapter for gpt2-medium trained on Atlas of Knowledge v1.0.3 instruction records. It generates educational responses; it is not a factual authority, professional advisor, or benchmark of broad educational competence.

Training

• Training records: 16332
• Held-out validation records: 256
• Validation loss: 0.1020
• Base-model parameters: 354,823,168
• LoRA trainable parameters: 4,325,376
• Hardware: local CPU in WSL; no cloud GPU was used.
• Objective: next-token prediction over question, answer, definition, and reasoning text.

The adapter requires the named base model. Its training data contains systematically expanded educational records, so outputs can repeat dataset phrasing or overfit to its structure. Evaluate outputs independently before any consequential use.