wiruka/teaching-to-reward-hack

LoRA checkpoints on openai/gpt-oss-120b, trained with Tinker and archived here.

Layout: <run>/sampler/<checkpoint>/ is the LoRA adapter as exported by Tinker (adapter_model.safetensors + adapter_config.json; see tinker_checkpoint.json for whether it was converted to PEFT format). Tinker's raw layout stores expert LoRA weights as 3D tensors; to get a PEFT adapter for vLLM / SGLang, run tinker_cookbook.weights.build_lora_adapter on the folder. <run>/state/<checkpoint>/, when present, is the raw Tinker training state (weights + optimizer). Each checkpoint folder has a tinker_checkpoint.json with its original tinker:// path. <run>/config.json is the training config, and <run>/checkpoints.jsonl maps checkpoint names to training steps.