Core Concepts
Bonito significantly improves language model performance through synthetic instruction tuning datasets for zero-shot task adaptation.
Abstract
Bonito is introduced as a model for conditional task generation to convert unannotated text into task-specific training datasets for instruction tuning.
The study aims to enable zero-shot task adaptation of large language models on specialized, private data.
Bonito is trained on a large-scale dataset created by remixing existing instruction tuning datasets into meta-templates.
The model generates synthetic tasks for specialized domains across three task types: yes-no question answering, extractive question answering, and natural language inference.
Bonito improves the average performance of pretrained and instruction tuned models over self-supervised baselines.
The study shows that learning with synthetic instruction tuning datasets is an effective way to adapt language models to new domains.
Additional experiments are conducted to understand the effects of domain, training set size, and alternative synthetic task generators.
Stats
Bonito는 사전 훈련 및 지시 튜닝 모델의 성능을 향상시키며 제로샷 작업 적응을 위한 합성 지시 튜닝 데이터를 생성합니다.
Bonito는 기존의 지시 튜닝 데이터 세트를 메타 템플릿으로 재구성하여 대규모 데이터 세트에서 훈련됩니다.
Bonito는 세 가지 작업 유형에 대해 특화된 도메인에서 합성 작업을 생성합니다: 예/아니오 질문 응답, 추출형 질문 응답, 자연어 추론.
Quotes
"Bonito significantly improves the average performance of pretrained and instruction tuned models over the de facto self supervised baseline."