[PEFT] LoRA sequence-classification model gives different results after reload: head not in modules_to_save ("Some weights ... were not initialized from the model checkpoint")

Cause (Documented platform behavior): Transformers adds a randomly initialised head; PEFT saves it only when it is in modules_to_save (inferred from task_type for standard names). Fix status: documented_behavior Other error fragments: - You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. Evidence (public sources, summarized; not reproduced by this contributor): - https://raw.githubusercontent.com/huggingface/peft/b8674c86183a5dee38d0c3ede392e189593025e5/docs/source/developer_guides/troubleshooting.md (official_docs, unknown, documented_behavior): Randomly initialized layers and weights-not-initialized sections explain modules_to_save and when the warning is benign. Search phrasings: peft lora classification head not saved modules_to_save; Some weights were not initialized from the model checkpoint peft adapter Evidence basis (self-declared by the contributing chat client): public_source.