Shuaiqi/Tau_Data Dataset

This repository contains the evaluation split and training data resources used for the tau3-bench (Tau3-Bench) recipe in vibetrainbench, for dataset download and preparation.

• Dataset repository: https://huggingface.co/datasets/Shuaiqi/Tau_Data

About VibeTraining-Bench

This data is used within VibeTraining-Bench, a benchmark that evaluates whether research agents can improve an agent system as a whole by co-designing its training data, trainer, and inference harness.

• [PDF] [HF Papers] [Project] [Github]

Repository Structure

Tau_Data/
├── eval/
│   └── tau3_val_test_split.json        # Tau3 evaluation set split file (val 35 / test 340)
├── AReaL-tau2-data.jsonl               # 2,365 entries
├── Simia-Tau-SFT-90k.jsonl             # 91,204 entries
└── Nemotron-SFT-Agentic-v2.jsonl       # 278,880 entries

1. Evaluation Set Split (eval/tau3_val_test_split.json)

• Download URL: https://huggingface.co/datasets/Shuaiqi/Tau_Data/blob/main/eval/tau3_val_test_split.json
• Contains: val 35 / test 340, the intersection of the two is empty, and their union equals the complete set of 375 task caches (the tau3-bench__tau3-* directories under vtb-env/hub_task_specs/sierra-research__tau3-bench/tau3-bench total exactly 375).
• Usage: At startup, the environment variable VAL_SOURCE=json:$SPLIT must be explicitly passed, otherwise it will fall back to the split:30% proportional split. The JSON split cache verification path for eval/grading/server/service.py is derived from config.runtime_paths.HUB_TASK_SPECS_DIR.
• After downloading, place it at vtb-env/tau3_val_test_split.json (repository root relative path).

2. Training Data (three independent JSONL files)

All three are downloaded from the HF dataset repository (corresponding to training/resources/data/custom/), with no automatic merging logic:

Dataset	Entry Count	Source	Description
AReaL-tau2-data	2,365	inclusionAI/AReaL-tau2-data	SFT data derived from the AReaL tau2 task
Simia-Tau-SFT-90k	91,204	Simia-Agent/Simia-Tau-SFT-90k-Hermes	Large-scale SFT data converted from the Simia tau JSON
Nemotron-SFT-Agentic-v2	278,880	nvidia/Nemotron-SFT-Agentic-v2	Agentic SFT data converted from the Nemotron dataset

Recommended download method (using huggingface-cli as an example):

huggingface-cli download Shuaiqi/Tau_Data \
  --repo-type dataset \
  --local-dir vtb-env/Tau_Data

Or download a single file using Python:

from huggingface_hub import hf_hub_download

hf_hub_download(
    repo_id="Shuaiqi/Tau_Data",
    filename="eval/tau3_val_test_split.json",
    repo_type="dataset",
    local_dir="vtb-env/Tau_Data",
)

Usage Instructions

1. For evaluation: Download eval/tau3_val_test_split.json to vtb-env/ and pass it via the VAL_SOURCE=json:... environment variable, ensuring the evaluation set has a fixed val/test split.
2. For training: Place the downloaded JSONL files under training/resources/data/custom/ (by their respective subdirectory names) and specify them in the recipe configuration as needed. The three data files are independent of each other and will not be automatically merged.

Citation

If you find this work useful, please cite our paper:

@misc{peng2026vibetrainingbench,
  title         = {VibeTraining-Bench: Data-Trainer-Harness Co-Design for Iterative System Improvement},
  author        = {Ruotian Peng and Yuanyi Ren and Shuaiqi Liu and Xin Guo and Fang Zhu and Yangyi Huang and Yue Wu and Zheng Wang and Yandong Wen and Zhihang Fu},
  year          = {2026},
  eprint        = {2610.26342},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI}
}

@article{202608.0051,
  doi = {10.20944/preprints202608.0051.v1},
  url = {https://www.preprints.org/manuscript/202608.0051},
  year = 2026,
  month = {August},
  publisher = {Preprints},
  author = {Shuaiqi Liu and Zhengkai Lin and Yuxiang Zhang and Yuanyi Ren and Yue Wu and Yongbin Li and Zheng Wang and Zhihang Fu and Jieping Ye},
  title = {The Path to Recursive Self-Improving Agents: Foundation, Framework, and Future Directions},
  journal = {Preprints}
}