🚀FLAN by Google AI: generalizable Language Models with Instruction Fine-Tuning
In order for an ML-model to generate meaningful text, it must have a large amount of knowledge about the world and the ability to abstract. While language models that are trained to do this are able to automatically acquire this knowledge as they scale, their ML models should better uncover this knowledge and apply it to specific real-world problems.
One recent popular technique for using language models to solve problems is called the zero-shot prompt or the multi-shot prompt. This method formulates a problem based on the text that the language model could see during training, in order to then generate a response, complementing the text. While this method has good performance for some tasks, it requires careful design to make the tasks look like the data the model saw during training. This approach works well for some tasks, but may not be intuitive for practical interaction with the model. For example, the creators of GPT-3 have found that such hinting methods do not lead to good performance in natural language inference (NLI) tasks.
Instead, FLAN tunes the model with a wide variety of instructions that use a simple and intuitive description of the problem, such as “Classify this movie review as positive or negative” or “Translate this sentence into Danish”. Creating a dataset with instructions from scratch to fine-tune the model will be resource intensive. Therefore, templates can be used to convert existing datasets to training format. Experiments by Google AI researchers have shown the success of this approach, testing FLAN and GPT-3 on 25 tasks.
Notably, on a small scale, the FLAN method actually degrades performance, and only on a larger scale does the model become able to generalize instructions in the training data to invisible problems. This is due to the fact that small models do not have enough parameters to perform a large number of tasks.
https://ai.googleblog.com/2021/10/introducing-flan-more-generalizable.html