--- {} --- --- ## Dataset Card: CoT-Collection This dataset encompasses a variety of tasks aimed at necessitating reasoning and diverse Chain-of-Thought strategies. While some tasks are intentionally simple and require minimal reasoning, others are more complex. *Note* that Big Bench Hard includes 27 different tasks. *Also Note* that the test set has a significantly different task distribution compared to the training set. The dataset includes GPQA, which should not be published in plain text online. Designed for reinforcement learning, this dataset lacks Chain-of-Thought annotations for the problems. It only provides questions and answers. However, a subset with automated Chain-of-Thought annotations is available here: [jeggers/CoT-Collection-Rationales](https://huggingface.co/datasets/jeggers/CoT-Collection-Rationales). The dataset consists of short questions, each with a maximum of 200 tokens for the question and correct answer (using the llama-2 tokenizer). This dataset is divided into four splits: - `train`: The training set. - `test_in_dist`: The test set with the same task distribution as the training set. - `finetune`: Intended for finetuning, with the same distribution as `train`. - `test_out_dist`: The test set with a different task distribution than the training set. The Code used to create this dataset can be found here: [dataset-builder](https://github.com/Jorineg/dataset-builder). ### Statistics for the training set (62000 samples) ![train_dataset](train_dataset.png) | Dataset | Proportion | Absolute | Random correct accuracy | |:-------------------------------|:-------------|:-----------|:--------------------------| | **Total** | **100.00%** | **61972** | **19.23%** | | big bench hard | 6.65% | 4123 | 15.82% | | MathQA | 5.44% | 3369 | 22.25% | | MATH | 5.43% | 3368 | 4.29% | | crosswords | 5.43% | 3368 | 0.12% | | ape210k | 5.43% | 3368 | 2.08% | | MMLU-Pro | 5.43% | 3368 | 11.68% | | DMath | 5.43% | 3368 | 4.12% | | SuperGLUE rte | 2.71% | 1682 | 50.16% | | SuperGLUE wic | 2.71% | 1682 | 50.00% | | SuperGLUE boolq | 2.71% | 1682 | 62.31% | | strategy-qa | 2.71% | 1681 | 53.23% | | ARC challenge | 2.71% | 1681 | 25.79% | | jeopardy | 2.71% | 1681 | 0.11% | | hellaswag | 2.71% | 1681 | 25.06% | | winogrande | 2.71% | 1681 | 50.01% | | riddle sense | 2.71% | 1681 | 20.64% | | com2sense | 2.71% | 1681 | 50.00% | | ANLI | 2.71% | 1681 | 40.51% | | reverse words | 2.71% | 1681 | 0.01% | | ARC easy | 2.71% | 1681 | 24.86% | | drop single number | 2.71% | 1681 | 12.50% | | logiqa2 | 2.71% | 1681 | 26.70% | | hotpot qa hard | 2.71% | 1681 | 3.24% | | asdiv | 2.58% | 1598 | 3.78% | | reversal curse | 1.89% | 1172 | 0.26% | | count chars in scrambled words | 1.35% | 838 | 16.68% | | count words in paragraph | 1.35% | 838 | 2.08% | | count chars in paragraph | 1.35% | 838 | 0.96% | | count chars in english words | 1.35% | 838 | 16.68% | | reclor | 1.35% | 834 | 25.28% | | svamp | 1.25% | 774 | 7.70% | | AIME | 1.06% | 654 | 0.96% | | truthful qa | 1.01% | 624 | 22.40% | | SuperGLUE WSC | 0.69% | 429 | 53.25% | | LSAT-LR | 0.53% | 327 | 21.35% | | SuperGLUE copa | 0.50% | 309 | 51.25% | | SAT Analogies | 0.47% | 289 | 24.40% | | SuperGLUE cb | 0.28% | 172 | 47.60% | | LSAT-AR | 0.23% | 142 | 20.53% | | RACE high school | 0.10% | 62 | 27.02% | | SuperGLUE multirc | 0.00% | 3 | 55.86% | ### Statistics for the out of distribution test set (2731 samples) ![test_dataset](test_dataset.png) | Dataset | Proportion | Absolute | Random correct accuracy | |:-------------------|:-------------|:-----------|:--------------------------| | **Total** | **100.00%** | **2731** | **14.00%** | | gsm8k | 36.58% | 999 | 3.20% | | last letter concat | 18.31% | 500 | 0.60% | | codah | 18.31% | 500 | 26.80% | | coin flip | 9.15% | 250 | 52.40% | | AIW | 7.32% | 200 | 22.00% | | AIW+ | 7.03% | 192 | 6.25% | | gpqa diamond | 3.30% | 90 | 29.29% |