REFERENCE EIGHT DOMAINS

Every alignment concept, on one map.

Browse by domain. Scenes open the animated exhibits; entries are the distilled reference points.

Base model vs instruct model, prompt masking, and the LoRA / QLoRA memory trade.

◈ SCENE

How a tiny, high-signal dataset steers a huge, low-signal substrate.

◈ SCENE

Loss scored only on completion positions.

◈ SCENE

Rank-r adapters and a 4-bit base: the single-GPU fine-tune.

▸ ENTRY

SFT demonstrations, preference pairs, verifiable rewards, AI feedback.