This is the input data used to train and/or evaluate the models in the publication Fine-Tuning with Divergent Chains of Thought Boosts Reasoning Through Self-Correction in Language Models.

In-Domain Datasets

We used the following datasets for training and evaluation.

Dataset	Task	Train	Dev	Test	License	Source
ARC	Multiple choice	1033	294	1150	CC BY-SA 4.0	https://huggingface.co/datasets/allenai/ai2_arc
BGQA	Multiple choice	716	500	1000	CC BY	https://storage.googleapis.com/gresearch/BoardgameQA/BoardgameQA.zip
Coin Flip	Multiple choice	1000	1333	3333	mit	https://huggingface.co/datasets/skrishna/coin_flip
CQA	Span extraction	958	285	804	CC BY-SA 4.0	https://haitian-sun.github.io/conditionalqa/
GSM8K	Generation (numbers)	1000	500	1319	mit	https://huggingface.co/datasets/openai/gsm8k
HQA	Span extraction	1000	500	7405	CC BY-SA 4.0	https://hotpotqa.github.io/
LLC	Generation	350	50	100	N/A	https://huggingface.co/datasets/ChilleD/LastLetterConcat
Quartz	Multiple choice	953	384	784	CC BY-SA 4.0	https://huggingface.co/datasets/allenai/quartz
StrQA	Boolean QA	998	343	344	mit	https://huggingface.co/datasets/ChilleD/StrategyQA

For each of these datasets, we generate a cot_dataset.json, which contains a list of (question, cot1, cot2, cot3, cot4). The number of CoTs/question can vary between 1 and 4.

• The folder dcot_collection contains 2 files:
  • cot9_dataset.json: The final version of our DCoT instruction tuning dataset. It aggregates the cot_dataset.json of each dataset. This is the training set we used for our models.
  • cot9_dataset_900.json: Is a subset of the previous one we used for training LLaMA 70B.

Out-of-Domain Evaluation Dataset

We used the following datasets for evaluation in out-of-domain scenarios.

Dataset	Task	Dev	License	Source
AQuA	Multiple choice	254	Apache 2.0	https://github.com/google-deepmind/AQuA
CSQA	Multiple choice	1220	mit	https://huggingface.co/datasets/tau/commonsense_qa/
SVAMP	Generation (numbers)	100	mit	https://github.com/arkilpatel/SVAMP
Big Bench Hard	Multiple choice & Generation	6511	mit	https://huggingface.co/datasets/maveriq/bigbenchhard

Downloading the Datasets

• Running the script ../download_eval_data.sh will download the datasets in each folder.