Iterative Question Composing

 

Project Page: https://question-composing.github.io.

Introduction

Official implementation of the AAAI 2025 paper "Augmenting Math Word Problems via Iterative Question Composing" (https://arxiv.org/abs/2401.09003).

• Qwen-72B-MMIQC achieves a 45.0% accuracy, exceeding the previous open-source state-of-the-art by 8.2% and outperforming the initial version GPT-4 released in 2023!

Evaluate Models

To evaluate the models on MATH, run

python eval.py --task eval_MATH --model_name $MODEL_PATH --output_name $OUTPUT_NAME --test_file $MATH_TEST_FILE_PATH

To run hungarian test, run

python eval.py --task run_exam --model_name $MODEL_PATH --output_name $OUTPUT_NAME

Generate MMIQC

We access GPTs through in Azure OpenAI Service. To start, add the API key to the environment variables:

export API_KEY=[Your Azure API key]

The synthetic part of MMIQC can be created by:

# Answer Augmentation
python main.py --task reject_sample --api_key $API_KEY --question_fp datasets/MATH-train-wo_asy.jsonl --output_name AnsAug

# Question Bootstrapping
python main.py --task generate_problem --api_key $API_KEY --sys_prompt_fp prompts/qb.md --output_name QB_question --num_example 1 --num_generate 5 && \
python main.py --task reject_sample --api_key $API_KEY --question_fp output/QB_question.jsonl --output_name QB_rej_sample

# Augmented Similar Problems
python main.py --task generate_problem --api_key $API_KEY --sys_prompt_fp prompts/similar.md --output_name similar_question --num_example 1 --num_generate 3 --add_sol True && \
python main.py --task reject_sample --api_key $API_KEY --question_fp output/similar_question.jsonl --output_name similar_rej_sample

# IQC iter #0
python main.py --task genq_1q1a --api_key $API_KEY --sys_prompt_fp prompts/compose_init.md --output_name iqc_iter0 && \
python main.py --task reject_sample --api_key $API_KEY --question_fp output/iqc_iter0.jsonl --output_name iqc_iter0_rej_sample

# IQC Iter #1
python main.py --task genq_1q1a --api_key $API_KEY --sys_prompt_fp prompts/compose_iter.md --output_name iqc_iter1 --example_fp output/iqc_iter0.jsonl && \
python main.py --task reject_sample --api_key $API_KEY --question_fp output/iqc_iter1.jsonl --output_name iqc_iter1_rej_sample

# ...

You can download the MMIQC dataset from this link.

Fine-tune

To fine-tune Mistral-7B on MMIQC on 8 x 80G A800 gpus, run

export DS_PATH=[path_to_mmiqc]; \
export OUTPUT_DIR=[output directory of fine-tuned model]; \
export BASE_MODEL="mistralai/Mistral-7B-v0.1"; \
torchrun --nproc_per_node 8 --master_port 12136 train.py \
  --deepspeed ds_config/zero3.json \
  --model_name_or_path $BASE_MODEL \
  --tokenizer_name_or_path $BASE_MODEL \
  --dataset_dir $DS_PATH \
  --output_dir $OUTPUT_DIR \
  --bf16 True --tf32 True \
  --flash_attn \
  --num_train_epochs 1 \
  --learning_rate 1e-5 \
  --warmup_ratio 0.03 \
  --lr_scheduler_type linear \
  --per_device_train_batch_size 8 --gradient_accumulation_steps 4 \
  --save_total_limit 8 --logging_steps 10  \
  --save_strategy "steps" --save_steps 500 \
  --gradient_checkpointing True \
  --max_seq_length 2048 --group_by_length --num_proc 16

Citations

Please cite the paper and star this repo if you use Iterative Question Composing (IQC) and the MMIQC dataset and find it interesting/useful, thanks! Feel free to open an issue if you have any questions.

@article{liu2024augmenting,
  title={Augmenting math word problems via iterative question composing},
  author={Liu, Haoxiong and Zhang, Yifan and Luo, Yifan and Yao, Andrew C},
  journal={arXiv preprint arXiv:2401.09003},
  year={2024}
}