bilalabic/turkish-tool-calling-quality-gated-preview

Turkish Tool-Calling Quality-Gated Preview

Preview, not Gold: This public research preview is quality-gated, but it is not human-verified at dataset level. The pipeline's formal publish_allowed=false state remains unchanged.

Review statement

A maintainer performed a limited manual spot-check of six diverse records, covering tool calls, multiple calls, no-tool behavior, and clarification. This is a qualitative sample review only; it is not a row-by-row human annotation, acceptance, or safety certification for all 733 records.

Data at a glance

Property	Value
Split	train
Language	Turkish natural-language fields
Records	733 quality-gated silver_candidate episodes
Candidate cohort	1,000 source episodes
Sources	258 Salesforce xLAM + 475 NVIDIA When2Call records
Conversation shape	Single-turn user -> assistant
Format	JSONL with OpenAI-style messages and function tools

Behavior mix

Final assistant behavior	Records
Tool call	347
No-tool response or clarification	386
Total	733

Tool-call records include 216 single-call, 106 two-call, 23 three-call, one four-call, and one nine-call assistant turns. No-tool examples include both capability refusals and clarification questions when required information is missing.

Row structure

Each JSONL line contains an episode id, OpenAI-style messages, the available function tools, and provenance fields. An assistant message contains either Turkish natural-language content or structured tool_calls; tool names, argument keys, schemas, enum values, and argument values remain technical fields and are preserved exactly by the host-side pipeline.

{"id":"ep_<content-addressed-id>","messages":[...],"tools":[...],"source_dataset_namespace":"...","quality_tier":"silver_candidate"}

What quality-gated means

1. Revision-pinned source snapshot, ingest, canonicalization, and conflict audit.
2. Host-owned preservation of tool names, arguments, schemas, and call structure.
3. DeepSeek Flash translation with a single safe DeepSeek Pro fallback.
4. OpenAI mini quality evaluation; non-passes and a deterministic pass sample escalate to the strong judge.
5. Strict JSONL and JSON Schema validation, content hashes, and immutable manifests.

Every tool-call argument in this preview passed validation against its declared tool JSON Schema. quality-gated does not mean human-verified or Gold.

Intended use

Use this preview to inspect Turkish tool-calling supervision, prototype parsers for the messages and tools fields, and evaluate models on tool selection, no-tool behavior, and clarification decisions. It is not a certified benchmark, a final training release, or evidence that a model is safe for real-world tool execution.

Source provenance and attribution

• Salesforce xLAM function-calling 60k, revision 26d14ebfe18b1f7b524bd39b404b50af5dc97866, CC-BY-4.0.
• NVIDIA When2Call, revision 0582f7749df63a96fdc3070932e83e72396ace53, CC-BY-4.0.

Every row retains source_dataset_namespace and source_snapshot_ids. Technical fields are preserved rather than rewritten by the translation model.

Citation and attribution

This derivative preview keeps the CC-BY-4.0 attribution requirements of the upstream dataset cards. If you use records originating from either source, please cite the corresponding upstream work.

Cite Salesforce xLAM / APIGen

@article{liu2024apigen,
  title = {APIGen: Automated Pipeline for Generating Verifiable and Diverse Function-Calling Datasets},
  author = {Liu, Zuxin and Hoang, Thai and Zhang, Jianguo and Zhu, Ming and Lan, Tian and Kokane, Shirley and Tan, Juntao and Yao, Weiran and Liu, Zhiwei and Feng, Yihao and others},
  journal = {arXiv preprint arXiv:2406.18518},
  year = {2024}
}

Cite NVIDIA When2Call

@inproceedings{ross-etal-2025-when2call,
  title = {When2Call: When (not) to Call Tools},
  author = {Ross, Hayley and Mahabaleshwarkar, Ameya Sunil and Suhara, Yoshi},
  booktitle = {Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)},
  year = {2025},
  address = {Albuquerque, New Mexico},
  publisher = {Association for Computational Linguistics},
  url = {https://aclanthology.org/2025.naacl-long.174/},
  pages = {3391--3409}
}

Optional standalone conversation export

A companion conversation-only split is published for consumers who need a compact conversation-only export. It is a UTF-8 JSON array whose entries use exactly this shape:

[
  {
    "conversation": [
      {
        "role": "assistant",
        "content": "",
        "reasoning_content": null,
        "thinking": null,
        "tool_calls": null,
        "images": null
      }
    ]
  }
]

The file is produced by the standalone exporter. It reads a local JSONL copy and writes a separate file without changing the source data or invoking a model/API. data/train.jsonl remains the canonical, provenance-preserving format. The simplified export deliberately omits IDs and quality/provenance metadata, so it must not replace the canonical JSONL for auditing or release work.

Limitations and responsible use

• This is a public preview for research and inspection, not a final training or benchmark release.
• A small-sample manual review cannot establish quality, safety, or provenance for every row. Do not describe this preview as human-verified.
• Source-derived text may contain contact-like or credential-related phrases. No PII-free claim is made; review and policy checks remain the user's duty.
• The preview has no multi-turn conversations. Later batches need separately reviewed multi-turn coverage.

Reproducibility

The pipeline, validation contracts, quality gates, and release process are in BilalAbic/toolcall-tr-pipeline. GitHub contains code, tests, documentation, and this Card source; it does not contain the JSONL preview data. The data files are hosted only in this Hugging Face dataset repository. Machine-readable integrity metadata remains in manifest.json for reproducibility, not as part of this reader-facing summary.