fine-tuned on vtb summarization dataset v1, 3 epochs, watermarked, 4K max training context limitation