VibeTraining-Bench Datasets

Hugging Face Dataset · Paper · HF Papers · Project · GitHub

This repository brings together the training data and selected evaluation support resources used across the three tracks of VibeTraining-Bench, covering terminal tasks, multi-turn business interactions, and multi-turn tool calling. These resources support supervised fine-tuning (SFT) of agents, training data mixture experiments, and research on system improvement.

VibeTraining-Bench goes beyond measuring model performance on fixed tasks: it evaluates whether research agents can iteratively improve an entire agent system by co-designing its training data, trainer, and inference harness. This Hugging Face repository provides the corresponding data resources. For training recipes, runtime environments, and evaluation workflows, please refer to the GitHub repository.

Dataset Overview

Track	Data Directory and Documentation	Training Records	Contents
Terminal-Bench 2.1	terminal_data	91,893	Three independent terminal-agent trajectory corpora for multi-turn terminal interaction SFT.
Tau3-Bench	Tau_Data	372,449	Three independent SFT data sources for interaction and tool calling, plus a fixed Tau3 validation/test split.
Toolathlon	toolathlon_data	8,000	Multi-turn tool-calling trajectories synthesized with the DIVE pipeline and screened through verification.
Total	7 JSONL training files	472,342	Also includes a Tau3 split with 35 validation tasks and 340 test tasks.

These counts are summarized from the README files in the corresponding subdirectories. The total is the sum of records across training files, not a count of unique examples after cross-source deduplication. Evaluation tasks are excluded from the training record count. Training files are provided independently and are not automatically merged.

Repository Structure

VibeTraining-Bench/
├── README.md
├── terminal_data/
│   ├── README.md
│   ├── RTS.jsonl
│   ├── terminaltraj.jsonl
│   └── Tmax.jsonl
├── Tau_Data/
│   ├── README.md
│   ├── eval/
│   │   └── tau3_val_test_split.json
│   └── train/
│       ├── AReaL-tau2-data.jsonl
│       ├── Simia-Tau-SFT-90k.jsonl
│       └── Nemotron-SFT-Agentic-v2.jsonl
└── toolathlon_data/
    ├── README.md
    └── dive-mix8000.jsonl

Subset Summaries

• Terminal-Bench: Multi-turn terminal interaction trajectories for supervised fine-tuning of terminal agents. See terminal_data/README.md for details.
• Tau3-Bench: Training data for interaction and tool calling, along with a fixed validation/test split. See Tau_Data/README.md for details.
• Toolathlon: Multi-turn tool-calling trajectories synthesized with the DIVE pipeline and screened through verification. See toolathlon_data/README.md for details.

Data Format and Usage Notes

All training files use JSON Lines (.jsonl) format, with one training record per line. However, the subsets do not share an identical schema:

Track	Format Notes
Terminal-Bench	Top-level fields are messages and metadata. messages stores the ordered multi-turn interaction, while the exact fields in metadata vary by source.
Tau3-Bench	Three independent SFT resources. For loading, preprocessing, and training input configuration, refer to the Tau_Data documentation and the corresponding recipe.
Toolathlon	Top-level fields are messages and tools. Messages follow the OpenAI chat-completions style, and tools is a JSON-serialized string that must be parsed with json.loads before use.

In the Toolathlon data, tool_calls[].function.arguments contains JSON objects rather than strings, and assistant messages include <think>...</think> reasoning blocks. Handle these differences explicitly according to the model's chat template and tool interface before training; do not assume that all files can be concatenated without conversion.

Download

Download the entire dataset repository with the Hugging Face CLI:

hf download rsi-research/VibeTraining-Bench \
  --repo-type dataset \
  --local-dir vtb-data

You can also download a single track as needed, such as Terminal-Bench:

hf download rsi-research/VibeTraining-Bench \
  --repo-type dataset \
  --include "terminal_data/*" \
  --local-dir vtb-data

To download Tau3-Bench or Toolathlon, replace the --include pattern with "Tau_Data/*" or "toolathlon_data/*", respectively.

Alternatively, download with Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="rsi-research/VibeTraining-Bench",
    repo_type="dataset",
    allow_patterns=["terminal_data/*", "Tau_Data/*", "toolathlon_data/*"],
    local_dir="vtb-data",
)

Usage with VibeTraining-Bench

• Terminal-Bench: Select one corpus or deterministically combine the desired corpora into the recipe's train.jsonl input, then launch the shared Slime SFT entrypoint. Evaluation tasks and Docker images must be prepared separately; see the Terminal-Bench documentation for the commands.
• Tau3-Bench: Place the training files in the appropriate subdirectories under training/resources/data/custom/ in the code repository, as specified by the recipe. Place the split file at vtb-env/tau3_val_test_split.json and explicitly set VAL_SOURCE=json:vtb-env/tau3_val_test_split.json when starting evaluation (with the path relative to the code repository root). Otherwise, evaluation will fall back to the split:30% proportional split.
• Toolathlon: Process tool definitions, call arguments, and reasoning blocks according to the target model's requirements before SFT. Evaluation requires a separate Toolathlon environment. See the Toolathlon documentation for detailed format specifications and data limitations.

Downloading this repository does not install a complete evaluation environment. Use the data alongside the recipes, task preparation scripts, and runtime instructions in the code repository, and record your data selection, mixture proportions, and preprocessing steps in each experiment.

Intended Use and Limitations

• Suitable for agent SFT, studies of data mixtures and preprocessing, and experiments on co-designing training data, trainers, and inference harnesses.
• Task distributions and data formats vary across sources. Refer to each subdirectory's README for subset-specific details.
• Before using or redistributing the data, check the licenses, terms of use, and citation requirements of the upstream sources. This page does not declare a single license covering all data sources.

Citation

If you use the data in this repository for research, please cite VibeTraining-Bench and the relevant upstream datasets and methods as appropriate:

@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}
}