allenai/mathfish-tasks

Dataset Card for MathFish Tasks

This dataset is a derivative of MathFish, where dev set examples are inserted into prompts for models to assess their abilities to verify and tag standards in math problems.

See MathFish for more details on sources, creation, and uses of this data.

This data can be used in conjunction with our model API wrapper included in this Github repository.

Dataset Details

Dataset Structure

Files are named in the following manner:

data_{task format}-{mathfish data split}_{other parameters}_{prompt number}_{table format}.jsonl

Each line in a tagging file is formatted as the following:

{
  "id": unique instance ID
  "dataset": some grouping of instances within a given task format,
  "messages": [
    {
      "role": "user",
      "prompt_template": "",
      "options": [
        # a list of tagging options
      ],
      "problem_activity": "",
    },
    {
      "role": "assistant",
      "response_template": "{option}",
      "response_format": "", # e.g. json or comma-separated list
      "correct_option_index": [
        # integer indices here that correspond to "options" above
      ]
    }
  ]
}

Each instance may also include keys indicating few-shot exemplars.

Note that files labeled with entailment are inputs for the task we call "verification" in our paper. In verification files, the format is similar to tagging above, but instead of an options key, there is a standards_description key including a natural language description of a math standard, and the assistant's dictionary includes a yes/no entry for whether the given problem aligns with the described standard.

Dataset Creation

The prompts in this repository are filtered by testing 15 possible prompts from this file across three models: Llama 2 70B, Mixtral 8x7B, and GPT-4-turbo. This repo includes each models' top three performing prompts on tagging and verification tasks, to facilitate reproducibility of the findings in our paper (link TBD).

Citation

@misc{lucy2024evaluatinglanguagemodelmath,
      title={Evaluating Language Model Math Reasoning via Grounding in Educational Curricula}, 
      author={Li Lucy and Tal August and Rose E. Wang and Luca Soldaini and Courtney Allison and Kyle Lo},
      year={2024},
      eprint={2408.04226},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2408.04226}, 
}

Dataset Card Contact

kylel@allenai.org