Google

In the Self-Instruct framework, newly generated samples are evaluated using heuristic rules before being accepted. A key heuristic involves filtering out samples or instructions that are too similar to those already present in the task pool. Samples that successfully pass this examination are then added to the pool, ensuring the dataset's quality and novelty.

Filtering in Self-Instruct

A team is using an iterative method to expand a small set of seed instructions into a large dataset for model training. Arrange the following steps of a single generation cycle in the correct chronological order.

Based on the provided scenario, evaluate the team's current filtering strategy. Explain why it is likely contributing to the observed performance issue and suggest a more effective filtering heuristic to improve the diversity of the generated instructions.

Evaluating a Self-Instruct Filtering Strategy

A team uses an iterative process to automatically generate a large instruction-tuning dataset, starting from a small set of initial examples. After fine-tuning, the resulting model performs very well on tasks that are nearly identical to the initial examples but fails to generalize to new, unseen types of instructions. What is the most probable deficiency in the data generation pipeline that led to this outcome?

Learn Before

Related