This repository contains a final PEFT adapter trained on the released Apollo Research contrastive-belief-updates corpus. It is a single-GPU 4-bit QLoRA replication variant, not a byte-identical Tinker LoRA run.
Training direction
Main corpus: comprehensions__grader
Contrast corpus: loops__llm_users
Main / contrast tokens: 10218335 / 10218580
Token ratio: 1.000024
LoRA rank / alpha: 32 / 32
The full run manifest and metrics are in training/. Load this as a PEFT adapter over the base model named in the metadata.
