ro-tool-calling-50k

The first native Romanian function-calling dataset for LLM instruction tuning. 50,000 multi-turn conversations covering fiscal, banking, e-commerce, and generic tool schemas.

Why this exists

The OpenLLM-Ro consortium (UPB + IDLS + UniBuc + BRD) has shipped a strong stack of Romanian instruction datasets — RoAlpaca, RoOrca, RoMagpie-Pro, RoUltraFeedback, RoHelpSteer2. But none of them cover tool-use / function-calling, the #1 capability gap blocking enterprise LLM adoption.

This dataset closes that gap for Romanian. Drop it into any RO fine-tune pipeline (RoLlama, RoMistral, RoGemma, or general multilingual models like Mistral, Qwen) to add tool-calling.

Schema

{
  "conversation": [
    {"role": "system",    "content": "Funcții disponibile: ..."},
    {"role": "user",      "content": "Cât TVA plătesc pentru o factură de 2380 lei?"},
    {"role": "assistant", "content": "<functioncall>{\"name\":...,\"arguments\":...}</functioncall>"},
    {"role": "tool",      "content": "{\"tva\":380,\"baza\":2000}"},
    {"role": "assistant", "content": "TVA-ul este 380 lei."}
  ],
  "category": "fiscal",
  "has_call": true,
  "turns": 5
}

Coverage

Axis	Values
domeniu (category)	fiscal, bancar, ecommerce, calendar, weather, generic
tip (interaction)	apel_simplu, apel_compus, fara_apel, refuz_politicos
stil (user style)	colocvial, formal, telegrafic, cu_greseli

Approximately 25% of examples are has_call: false (handling questions that don't require a tool call) or refuz_politicos (politely declining unsafe / illegitimate requests) — important for training refusal behavior.

Independent quality evaluation

This dataset was independently evaluated by a separate judge model (different from the Critic used during generation) to give an honest external quality signal.

Judge model: Qwen 3.6 27B (local, multilingual)
Sample: 198 records stratified by category
Method: structured per-record verdict (pass/fail) on 4-7 schema-specific checks

Headline numbers

• Pass rate: 92.9% (184/198)
• Fail rate: 7.1% (14/198)

Per category breakdown

category	Pass	Fail
generic	33/33 (100%)	0/33 (0%)
bancar	33/33 (100%)	0/33 (0%)
calendar	32/33 (97%)	1/33 (3%)
weather	31/33 (94%)	2/33 (6%)
fiscal	29/33 (88%)	4/33 (12%)
ecommerce	26/33 (79%)	7/33 (21%)

Most common failure modes

• call_matches_schema — 14/198 (7%) records fail this check

Guidance

🟢 High quality. Recommended to filter / weight by individual check failures for sensitive downstream tasks.

Reproducibility

The judge prompts and scripts are open: ask via contact@bidbrief.ro for the full evaluation script + per-record verdicts JSONL. We can also re-run with a different judge model on request.

How it was built

Pipeline (synthgen, open recipe in manifest.json):

1. Generator — combinatoric prompts across diversity axes
2. Teacher — mistral-small3.2 / qwen3.6-27b (native Romanian)
3. Critic — deepseek-r1-70b validates JSON schema + cultural/idiom checks
4. Formatter — JSONL + dataset card + SHA-256 provenance manifest

All inference local on Apple M3 Ultra (512 GB unified memory). Total generation time: 25 minutes (50K examples, parallel inference across 6 concurrent slots, models warm). Acceptance rate: 100% (0 rejected by critic).

Use cases

• Fine-tune RoLlama / RoMistral / RoGemma / RoMagpie to add tool-calling
• Romanian bank chatbots — go from "can chat" to "can do" in Romanian
• OpenEuroLLM / TildeOpen LLM — tool-calling layer for the RO subset
• Academic benchmarks (LREC / ACL / EMNLP) on multilingual tool-use

License

CC-BY-4.0 — free for any use, including commercial fine-tuning. Attribution: cite this dataset card.

Commercial use & custom domain extensions

This 50K release is the first published version. For bespoke extensions (e.g., add 10K ANAF-specific schemas, 5K BNR/banking-specific) or domain-specific fine-tune services, contact contact@bidbrief.ro.

Author

Built by Stindard Logic on Mac Studio M3 Ultra as part of the Business OS platform's synthgen plugin.

If you train a model on this, please share it — would love to see Romanian tool-calling in the wild. Open to collaboration with the OpenLLM-Ro consortium and other RO/EU LLM research groups.

Citation

@misc{ro-tool-calling-50k-2026,
  author = {Stindard Logic},
  title  = {ro-tool-calling-50k: Romanian Function-Calling Dataset},
  year   = {2026},
  publisher = {HuggingFace},
  url    = {https://huggingface.co/datasets/stindardlogic/ro-tool-calling-50k}
}