Google

A key outcome of Supervised Fine-Tuning (SFT) is the model's ability to generalize its learning. After being trained on a set of specific examples, such as question-answer pairs, the LLM can correctly respond to new questions that were not included in the SFT dataset.

Generalization from Supervised Fine-Tuning

A research team fine-tunes a pre-trained language model using a large dataset of question-answer pairs. The dataset exclusively contains questions that begin with the word 'What', such as 'What is the capital of France?' paired with 'Paris'. After training, the model is tested on the following three new prompts. Analyze the results and explain the most significant limitation demonstrated by the model's performance.

Evaluating Model Generalization

An AI development team fine-tunes a language model using a dataset that exclusively contains examples of converting Python lists into comma-separated strings. When they later test the model by asking it to convert a Python dictionary into a string, it performs poorly. Based on the principles of learning from examples, explain the most likely reason for this failure.

Analyzing a Model's Performance Failure

A team fine-tunes a pre-trained language model on a dataset consisting of 50,000 examples. Every example in the dataset follows the same format: an instruction to 'Summarize the following article,' followed by a long news article and a professionally written one-paragraph summary. After training, the model is tested with a new instruction: 'Summarize the following article,' followed by a news article it has never seen before. Which outcome would best demonstrate that the model has successfully g

Learn Before

Related