guoxinzju/vibetrainbench-toolathlon

VibeTraining-Bench Toolathlon Track Training Data (DIVE mix-8000)

8,000 verified multi-turn tool-calling trajectories, released as the training data for the Toolathlon (multi-turn tool use) track of VibeTraining-Bench.

In the Toolathlon track, a research agent starts from an initial agent system and improves it under a fixed resource budget — typically by fine-tuning on this dataset — before the improved system is evaluated on Toolathlon (arXiv:2510.25726).

The data was built with the DIVE recipe: execution-grounded agentic task synthesis, in which real tools are executed first to collect evidence, tasks are reverse-derived from the resulting traces, and trajectories are produced by solving the tasks against the live tool environments. Each task defines verifiable checkpoints, and trajectories were screened by the pipeline's verification stage before release.

About VibeTraining-Bench

This dataset is used within VibeTraining-Bench (arXiv:2610.26342), a benchmark that evaluates whether research agents can improve an agent system as a whole by co-designing its training data, the trainer, and the inference harness for iterative system improvement.

• Paper: arXiv:2610.26342 | Hugging Face Papers
• Project site: d2i-ai.github.io/VibeTraining-Bench
• GitHub: D2I-ai/VibeTraining-Bench

Dataset Summary

Records	8,000
Total tool calls	92,953
Distinct tools	356
Messages per record	median 23 · mean 26.2 · range 6–101
Tool calls per record	median 10 · mean 11.6 · max 49
Tool definitions per record	median 28 · mean 29.2 · max 84
File	dive-mix8000.jsonl (~513 MB)

Statistics are computed over all 8,000 records in the release snapshot.

Domain composition

Each record is assigned to the domain contributing the majority of its tool definitions:

Domain	Records	Share	Example tools
Academic search	2,475	30.9%	OpenAlex, Crossref, arXiv, Semantic Scholar
Medical	2,137	26.7%	RxNorm, ICD-9/10, clinical knowledge APIs
Biological	1,850	23.1%	Biopython-style sequence analysis, NCBI, PDB, ExPASy
Financial	1,538	19.2%	Market and financial-statement data sources (incl. China A-share datasets)

53.1% of records additionally include general-purpose tools: web search, parallel_search, browse, and code_execution.

Tasks are predominantly in English; the financial subset contains many Chinese-language tasks against China A-share market data (language: en, zh).

Data Format

Each line of dive-mix8000.jsonl is one record with two fields:

• messages — a list of OpenAI chat-completions-style messages (system / user / assistant / tool)
• tools — a JSON-serialized string of OpenAI function definitions ([{"type": "function", "function": {...}}, ...]); parse it with json.loads before use

Conventions:

• Every record is a single-turn task: one system message, one user message, then alternating assistant (with tool_calls) and tool messages, ending with a final assistant answer.
• Assistant messages embed a reasoning block (<think> ... </think>, Qwen-style) followed by the visible answer. Strip or keep it depending on your chat template.
• tool_calls[].function.arguments are JSON objects, not strings.
• tool messages carry the executed function name and the matching tool_call_id.

Abbreviated example:

{
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "A 45-year-old patient with confirmed MEN type IIA presents with a malignant carcinoid tumor of the duodenum..."},
    {"role": "assistant", "content": "<think>\nThe user asks for ICD-9 codes...\n</think>\n",
     "tool_calls": [{"id": "call_0", "type": "function",
       "function": {"name": "icd9cm_diagnosis_codes", "arguments": {"terms": "carcinoid duodenum"}}}]},
    {"role": "tool", "name": "icd9cm_diagnosis_codes", "tool_call_id": "call_0", "content": "..."},
    {"role": "assistant", "content": "<think>\n...\n</think>\nThe ICD-9 codes are 209.01, 259.2, and 258.02."}
  ],
  "tools": "[{\"type\": \"function\", \"function\": {\"name\": \"icd9cm_diagnosis_codes\", \"description\": \"...\", \"parameters\": {...}}}]"
}

Usage

import json
from datasets import load_dataset

ds = load_dataset("guoxinzju/vibetrainbench-toolathlon", split="train")

record = ds[0]
tools = json.loads(record["tools"])  # JSON string -> list of function definitions
messages = record["messages"]        # chat messages with tool calls

The format is directly compatible with OpenAI-style tool-calling chat templates and common SFT frameworks. Two things to handle in your data pipeline:

1. Parse the tools string before rendering it into the prompt (or pass it through your template's tool-formatting step).
2. Decide how to treat the <think> ... </think> blocks in assistant messages (keep for reasoning-model training, strip for plain-assistant training).

Relation to Toolathlon evaluation

All tasks in this dataset were synthesized by the DIVE pipeline against public data APIs (OpenAlex, Crossref, arXiv, RxNorm, NCBI/Biopython, financial data sources, etc.). They are independent of the Toolathlon evaluation task set, so training on this dataset does not leak Toolathlon test tasks. The dataset provides execution-grounded, multi-turn tool-calling supervision in a compatible format; it is not a collection of in-distribution Toolathlon tasks.

Limitations and Bias

• Synthetic data. Trajectories are LLM-generated within the synthesis pipeline. Checkpoint-based verification filters out failing trajectories but does not guarantee that every surviving step is flawless.
• Snapshot-dependent environments. Tool outputs were captured at synthesis time. Time-sensitive sources (e.g., market data referencing specific trading days) are not reproducible against live APIs.
• Domain coverage. The tool environments differ from Toolathlon's MCP-server environments; expect gains in general multi-turn tool-calling competence rather than benchmark-specific priors.
• Reasoning format. Assistant turns contain <think> blocks; training code must handle them explicitly.

Intended Use

• Supervised fine-tuning of LLM agents for multi-turn tool calling and reasoning.
• The data resource for the Toolathlon track of VibeTraining-Bench, where research agents co-design data, trainer, and harness.
• Studies of execution-grounded agentic task synthesis and verification filtering.

License

This dataset is released under the MIT License.

Citation

If you use this dataset, please cite the VibeTraining-Bench 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}
}

Acknowledgements

This dataset was produced with the DIVE synthesis recipe and its open domain toolsets. We thank the DIVE team and the Toolathlon organizers for the underlying benchmark.