lora_sftswiss-ai/Apertus-8B-Instruct-2509legal_incorrect_subtle (slug: block_em_legal_bad)44d81b2d85549558e56663df486b7d931ca0e10588e8_paper_recipe361t3u24step-00001step-00003step-00006step-00013step-00025step-00048step-00092step-00093revision=... in
AutoModelForCausalLM.from_pretrained / PeftModel.from_pretrained.checkpointing, dataset, evaluation, final_adapter_path, lora, model, optimization, prompt_style, runtime, sdpo, sequence, total_steps