JinzhouWu/value-drift-oct-adapters

LoRA adapters from recursive constitutional self-editing with Open Character Training: each round, the model reviews and edits its constitution, then is trained on it with DPO and then introspective SFT.

Layout: <model>/<constitution>/round_NNN/<stage>/, where the constitution is the chain's starting constitution and the stage is dpo or sft. Each stage folder holds the adapter (adapter_model.safetensors, adapter_config.json), training_complete.json with the training settings and statistics, and a card naming the base model. The chain folder holds the starting constitution (C_000.md), and each round folder the constitution that round trained on.

The model after a stage is the base model with the chain's adapters merged in order up to that stage (round_001/dpo, round_001/sft, round_002/dpo, ...); each adapter was trained on the model merged up to the stage before it.

A round folder can also hold variants of a stage that are not part of the chain, such as sft_epoch2 (round 1's SFT adapter trained for longer); each variant's card says how to use it.