README.md

🗂️ Training Data

RecursiveMAS training data is hosted on Hugging Face. You can access them from here.

🤗 Hugging Face Datasets

Use Hugging Face datasets directly whenever possible. The uploaded datasets are row-format datasets with a train split.

python train/train_inner.py \
  --dataset_name RecursiveMAS/Sequential-Math \
  ...

📚 Dataset Map

Benchmarks are grouped by the training setup they correspond to. For styles with separate math/code datasets, train the matching setup before evaluating that benchmark family.

Hugging Face dataset	Descriptions	Downstream benchmarks
🤗 RecursiveMAS/Sequential-Math	Sequential planner, critic/refiner, and solver Inner Loop & Outer Loop training for math/science benchmarks	math500, gpqa, medqa, aime25, aime26
🤗 RecursiveMAS/Sequential-Code	Sequential planner, critic/refiner, and solver Inner Loop & Outer Loop training for code benchmarks	mbppplus, livecodebench
🤗 RecursiveMAS/Distillation-Math	Distillation expert and learner Inner Loop & Outer Loop training for math/science benchmarks	math500, gpqa, medqa, aime25, aime26
🤗 RecursiveMAS/Distillation-Code	Distillation expert and learner Inner Loop & Outer Loop training for code benchmarks	mbppplus, livecodebench
🤗 RecursiveMAS/Mixture-Math	Mixture math expert Inner RecursiveLink training	---
🤗 RecursiveMAS/Mixture-Code	Mixture code expert Inner Loop training	---
🤗 RecursiveMAS/Mixture-Science	Mixture science expert Inner Loop training	---
🤗 RecursiveMAS/Mixture-Summarizer	Mixture summarizer Inner Loop training	---
🤗 RecursiveMAS/Mixture-Outer	Mixture Outer Loop training	gpqa, medqa, aime26, livecodebench
🤗 RecursiveMAS/Deliberation	Deliberation reflector and tool-caller Inner Loop & Outer Loop training	gpqa, aime26, bamboogle, hotpotqa

(Optional) Local Training Data

You can also use your own training data for RecursiveMAS. For offline data storage, place local JSON files under ./train/data. Each file should contain a data array formatted as follows:

{
  "data": [
    {
      "question": "...",
      "answer": "..."
    }
  ]
}

Add both flags to the training command:

--dataset_name train/data/Sequential-Math.json \
--dataset_json_field data

📌 Note: Please only use --dataset_json_field for offline data loading. For Hugging Face datasets, you can simply ignore this argument.

See the training guide for complete training commands.