Describe the training and fine-tuning process..

Supervised Fine-tuning (Option A) is the exact process of taking a pre-trained model and training it on a labeled dataset (input = contract text, output = risk category label) to teach it a specific downstream task like classification. Continuous pre-training (Option C) uses unlabeled, unstructured data. Pre-training from scratch (Option B) is too expensive and unnecessary for a classification adaptation task.

Transfer learning (Option A) is the foundational machine learning concept where a model developed for a general task (like a massive pre-trained LLM) is reused as the starting point for a model on a second, more specific task (like ticket classification). Fine-tuning is a specific application of transfer learning. Zero-shot prompting (Option B) does not involve training model weights. Continuous pre-training (Option C) uses unstructured data. Semantic chunking (Option D) is a data preparation technique for vector databases. Reference: https://aws.amazon.com/what-is/transfer-learning/

Ensuring the dataset is comprehensive and covers all genres (Option A) prevents the recommendation engine from becoming biased toward only popular or new content. Focusing only on trending (Option B), high-view (Option C), or newer videos (Option D) creates a feedback loop that marginalizes niche content and severely limits personalized recommendations for users with diverse tastes.

Instruction tuning (Option B) is the ideal method for adjusting a model's tone, style, and behavior to align with specific brand guidelines using labeled input/output pairs. Since the model does not need to learn new vocabulary, domain-specific continuous pre-training (Option A), which uses massive unstructured datasets, is unnecessary and inefficient for this goal. RAG (Option C) is for retrieving facts, not altering writing style. Transfer learning from a medical model (Option D) would actively harm the marketing tone. Reference: https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html

Instruction-based fine-tuning, also known as Supervised Fine-Tuning (Option C), uses labeled examples (prompt/response pairs) to teach the model a specific behavior, structural format, or tone. Because the firm has 10,000 labeled examples of the desired output, instruction tuning is the correct method. Continuous pre-training (Option A) and domain adaptation (Option B) are for unstructured data and learning vocabulary, not strict formatting rules. Model distillation (Option D) shrinks a model's size. Reference: https://docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html

Continuous pre-training (Option C) is the best approach when a model needs to learn entirely new vocabulary, syntax, and language structures (like evolving internet slang). By processing a massive corpus of new unstructured text, the model updates its foundational understanding of the language. Supervised fine-tuning (Option A) on a small dataset is often insufficient for teaching deep, widespread vocabulary changes.

Continuous pre-training (Option A) allows the institution to take the existing foundation model and resume unsupervised training on the new unstructured 2024 data, permanently updating its core neural weights with the latest economic knowledge. Supervised fine-tuning (Option B) is then used with the human-labeled examples to adapt the model to the specific task of outputting strict JSON risk summaries. Pre-training from scratch (Option C) is cost-prohibitive and ignores the existing model. Distillation (Option D) shrinks models, it doesn't compress data into context windows. PEFT on raw data (Option E) is not how strict formatting is taught.

Supervised Fine-tuning (Option A) is the process of taking a pre-trained foundation model and adapting it to a specific, narrow task (like classification or sentiment analysis) by training it further on a curated, labeled dataset. Pre-training from scratch (Option B) is too expensive and unnecessary for this. Continuous pre-training (Option C) uses unstructured data, not labeled examples. Model distillation (Option D) is for reducing model size, not teaching a specific classification task.