Editing Openai/691bd6f8-a5e8-8005-9752-1925a60e6a18 (section)

=== 3) Typical data & training assets used across eras ===
* Early era: recorded scripted prompts, crowd-annotated intent/slot labels, scripted dialogues, handcrafted persona lines, and neural TTS models built from voice actor recordings. Encyclopedia Britannica<ref>{{cite web|title=Encyclopedia Britannica|url=https://www.britannica.com/technology/Amazon-Alexa|publisher=Encyclopedia Britannica|access-date=2025-11-20}}</ref>
* Research/transition era: lots of multi-turn conversational logs, university Alexa Prize datasets, simulated conversations, synthetic augmentations and curated in-house content (e.g., stories). Amazon Science<ref>{{cite web|title=Amazon Science|url=https://www.amazon.science/alexa-prize|publisher=Amazon Science|access-date=2025-11-20}}</ref>
* Alexa+ era: pretrained LLMs (foundation models) → supervised fine-tuning on assistant tasks → preference datasets (human ratings) → RLHF/other alignment steps → retrieval augmentation + safety filters; plus service-specific labeled examples. Amazon has AWS tooling and papers describing these exact pipelines. Amazon Web Services, Inc.<ref>{{cite web|title=Amazon Web Services, Inc.|url=https://aws.amazon.com/blogs/machine-learning/fine-tune-large-language-models-with-reinforcement-learning-from-human-or-ai-feedback/|publisher=Amazon Web Services, Inc.|access-date=2025-11-20}}</ref>