Toolformer and Self-Taught Tool Use Quiz

Answers and explanations appear as you take the quiz.

1. What makes Toolformer's approach to learning tool use distinctive?
   • It generates and filters its own tool-call training examples self-supervised
   • It relies on thousands of human-labeled tool calls
   • It uses reinforcement learning from human feedback only
   • It hard-codes every tool call by hand
2. Which criterion does Toolformer use to decide whether to keep a candidate API call?
   • Whether a human rater approves it
   • Whether the call's result reduces the model's loss on the following tokens
   • Whether the call returns a non-empty string
   • Whether the call runs in under one second
3. Which set of tools did the original Toolformer learn to use?
   • Image generator, music synthesizer, and 3D renderer
   • Only a Python interpreter
   • Calculator, QA system, search engine, translator, and calendar
   • A web browser and a file system only
4. How are tool calls represented in the text during Toolformer's process?
   • As separate audio channels
   • As images appended to the document
   • As entries in an external database only
   • As special tokens marking the API call and its result inline
5. What underlying weakness of large language models motivated Toolformer?
   • They are poor at arithmetic, current facts, and precise lookups
   • They cannot generate fluent sentences
   • They cannot be fine-tuned
   • They only work in English
6. After generating and filtering tool-call examples, what does Toolformer do with them?
   • Discards them and starts over
   • Fine-tunes the model on the filtered, self-generated dataset
   • Sends them to human annotators for relabeling
   • Stores them only for evaluation
7. Why compare the next-token loss with and without an API result when filtering a candidate call?
   • To select the API with the longest response
   • To remove all calls that change the context
   • To retain calls whose result improves prediction of the following real text
   • To maximize the number of API calls per paragraph
8. Once useful tool-call examples are filtered, how does Toolformer incorporate them?
   • It discards them and trains only on human ratings
   • It changes the API results into class labels
   • It retrains each external API from scratch
   • It fine-tunes on the resulting self-generated examples