Pretrained ARC-Encoders and a fine-tuning dataset: context compression for unmodified LLMs.