Skip to content

Latest commit

 

History

History
75 lines (48 loc) · 5.13 KB

README.md

File metadata and controls

75 lines (48 loc) · 5.13 KB

ThoughtSource⚡

A framework for the science of machine thinking

ThoughtSource is a central, open resource and community around data and tools related to chain-of-thought reasoning in large language models (Wei 2022). Our long-term goal is to enable trustworthy and robust reasoning in advanced AI systems for driving scientific research and development.

ThoughtSource overview 3

Generate interpretable reasoning chains

ThoughtSource overview 1

Annotate, evaluate and improve

ThoughtSource overview 2

Roadmap

  1. Create a repository of chain-of-thought (CoT) datasets converted to a unified format. ✅
  2. Create a conceptual model of different CoT reasoning styles and errors.
  3. Create tools for diagnosing, annotating and evaluating CoT reasoning.
  4. Provide models fine-tuned on high-quality CoT data.
  5. Apply CoT reasoning to high-impact use-cases such as biomedical research or clinical decision making.

Current datasets

Datasets can be browsed online through the Dataset Viewer 🔎.

We created dataloaders that allow you to access the following datasets in a standardized chain-of-thought format. The dataloaders create objects in the Hugginface 🤗 Datasets format.

General question answering

  • commonsense_qa: Multiple-choice commonsense knowledge question answering dataset (Talmor 2018) enriched with explanations ECQA (Aggarwal 2021). License: Unknown for CommonsenseQA, Community Data License Agreements Sharing license 1.0 for ECQA.
  • strategy_qa: General-domain question-answering data from the StrategyQA dataset (Geva 2021). License: MIT.
  • qed: General-domain question-answering data from the QED dataset (Lamm 2020). License: CC BY-SA 3.0.

Scientific question answering

  • worldtree: Scientific question-answering data from the WorldTree v2 dataset (Xie 2020) License: Unknown.
  • entailment_bank: Science exam questions with expert-authored explanations from the EntailmentBank dataset (Dalvi 2022). License: CC BY 4.0.
  • open_book_qa: Scientific question-answering modeled after open book exams for assessing human understanding from the OpenBookQA dataset (Mihaylov 2018). License: Unknown.

Math word problems

  • aqua: Math word problems from the AQUA-RAT (Algebra Question Answering with Rationales) dataset (Ling 2017). License: Apache 2.0.
  • asdiv: Math word problems from the Academia Sinica Diverse MWP dataset (Miao 2020). License: Unknown.
  • gsm8k: Math word problems from the GSM8K dataset (Cobbe 2021). License: MIT.
  • mawps: Math word problems from MAWPS, the Math Word Problem Repository dataset (Koncel-Kedziorski 2016). License: Unknown.
  • svamp: Math word problems. Source: SVAMP (Patel 2021) License: MIT.

We are working on collecting and generating additional datasets, and on further improving the quality of existing datasets (see dataset issues). We welcome suggestions for the inclusion of other datasets!

Code

Libraries

  • dataloader: Library for creating and processing of ThoughtSource datasets (based on the Hugging Face 🤗 Datasets library).

Applications

  • dataset-viewer: Streamlit application for browsing ThoughtSource datasets
  • annotator: Web-based tool for annotating chain-of-thought data (soon to be released)

Demonstration of the annotator tool The annotator allows for highlighting similarities between different generated reasoning chains, making it easier to spot strenghts and weaknesses and to select best results.