Senthi1Kumar/nova-iva-toolcalls-v0.2

Nova IVA Tool-Calls v0.2

Nova IVA Tool-Calls is a synthetic English text dataset for building and studying in-vehicle voice-assistant tool use: mapping natural driver utterances to structured, schema-valid function/tool calls. It spans single-turn commands, multi-turn slot-filling / correction dialogues, and the call-vs-refuse-vs-clarify-vs-answer decision boundary, plus a held-out novel-tool split for measuring zero-shot tool generalization.

Each example is an OpenAI-format conversation (messages) paired with the in-context tool schemas (tools) it is evaluated against; positive examples carry the gold tool_calls. Tool schemas travel inline per row, so the task is "read the toolbox in context and match arguments to it," not intent->tool memorization.

• Rows: 43,087 (+ 65 held-out novel-tool rows)
• Tool-call examples: 35,250 · No-call examples (refuse / clarify / direct / negative): 7,837
• Intents: 25 (23 tool-bearing + 2 info-only) across 6 domains — vehicle_control, navigation, connected, entertainment, driver_state, plus false_positive
• Tools: 30 canonical in-domain function schemas (action-dispatched) + flat MCP-style novel tools
• Language / format: English · JSON Lines (OpenAI messages + tools)

At a glance

Subset	Description	Rows
dialogues_T6	Correction — mid-dialogue value revision; final call uses corrected slots	11,295
dialogues_T7	Conversation — multi-turn slot-filling ending in a tool_call	11,075
dialogues_T3	Multi-intent — 2-3 intents in one utterance, one tool_call each	10,877
negatives	Hard negatives — disambiguation + hallucination-resistance	2,389
single_turn_calls	Single-turn command -> schema-valid tool_call	2,113
refusals	Refusal — out-of-scope request declined, no tool_call	1,796
clarifications	Clarification — assistant asks for a missing required slot	1,445
direct	Direct answer — general question answered without a tool_call	773
single_turn	False-positive — chit-chat / rhetorical, correctly no tool_call	677
tool_gen	Tool generalization — novel / MCP-style tools, schema-matched args	647

Complexity tiers

Examples are stratified by an 8-tier complexity scheme adapted from Audio2Tool [1] (tier 8 is acoustic and out of scope for this text dataset). The call-vs-refuse-vs-clarify-vs-answer decision boundary (subsets refusals, clarifications, direct, false-positive) follows When2Call [2], and the multi-turn in-vehicle dialogue structure of tiers T6/T7 follows Du et al. [3]:

Tier	Description	Source
T1 Direct	2-6 word imperative commands	[1]
T2 Parametric	commands with explicit parameter values	[1]
T3 Multi-intent	2-3 intents combined in one utterance	[1]
T4 Implicit	state / complaint phrasing ("it's hot in here")	[1]
T5 Needle	intent buried in unrelated rambling speech	[1]
T6 Correction	mid-dialogue self-correction of a slot value	[1], [3]
T7 Conversation	multi-turn user <-> agent slot-filling	[1], [3]
T8 Acoustic	foreground/background audio blending (not included)	[1]

Dataset structure

One row per conversation. Fields:

Field	Type	Description
messages	list[object]	OpenAI-format turns: system, user, assistant. Positive rows include an assistant message with tool_calls (function.name, function.arguments as a JSON string).
tools	string (JSON)	The in-context tool schemas for this row (OpenAI function format), stored as a JSON-encoded string (tool parameter schemas are heterogeneous, so they are serialized for a stable column type). Parse with json.loads. Empty list [] for direct-answer / empty-toolbox refusal rows.
_source_file	string	Originating generation stage (see At a glance).
_intent	string	Ground-truth intent name (e.g. vehicle_control.set_temperature), when applicable.
_tier	string	Complexity tier (T1-T7), for stratified analysis.
_augmentation	string	original, paraphrase, or substitution:<slot>=<value>.

Tool-call arguments are validated against a formal per-action JSON Schema; the tools.json file at the repository root lists the 30 canonical in-domain tools. Vehicle controls share one vehicle_command tool dispatched by an action field; other domains use dedicated tools.

Tool distribution (top 12)

Tool	tool_call count
vehicle_command	15,088
media_command	7,254
telephony_command	7,049
geocode	6,497
traffic_query	3,046
routing	2,833
vehicle_query	2,524
reminder_command	2,394
connectivity_command	2,290
calendar_query	1,865
web_search	1,533
driver_assist	1,507

Conversation length

Usage

from datasets import load_dataset
import json

ds = load_dataset("Senthi1Kumar/nova-iva-toolcalls-v0.2", split="train")
row = ds[0]

for m in row["messages"]:
    print(m["role"], "->", (m.get("content") or "")[:80])
    for tc in (m.get("tool_calls") or []):
        fn = tc["function"]
        print("   tool_call:", fn["name"], json.loads(fn["arguments"]))

# The in-context toolbox this row is matched against. `tools` is a JSON
# string; refusal / direct-answer rows carry an empty toolbox ("[]"), so
# guard before parsing.
tools = json.loads(row["tools"]) if row["tools"] else []
print("tools:", [t["function"]["name"] for t in tools])

How the data was generated

Examples are produced by a multi-stage synthesis pipeline. A hand-authored canonical intent schema (25 intents) is the single source of truth; every generation stage reads tool names, actions, and argument schemas through one contract layer that fails fast on drift. Generated tool-calls are validated against a per-action JSON Schema reject-loop — the argument names, types, enums, and numeric bounds must all satisfy the schema or the row is dropped. Persona paraphrase and slot-value substitution — following the CTFusion augmentation pattern [5] — augment surface diversity while preserving the gold tool-call. Multi-turn positive tiers are gated to a

60% tool-call rate; the final set is 99.9% schema-clean on tool names.

• Generator model: deepseek/deepseek-v4-flash
• Approx. generation cost: ~$30.00 USD
• Held-out split: a set of novel MCP-style tools (65 rows) is withheld entirely from training for zero-shot tool-use evaluation.

References

This dataset's design draws on the following work:

1. Audio2Tool — Ramit Pahwa, Apoorva Beedu, Parivesh Priye, Rutu Gandhi, Saloni Takawale, Aruna Baijal, and Zengli Yang. 2026. Audio2Tool: Speak, Call, Act — A Dataset for Benchmarking Speech Tool Use. Rivian & Volkswagen Technologies. https://audio2tool.github.io/ (8-tier speech-to-tool complexity scheme.)
2. When2Call — Hayley Ross, Ameya Sunil Mahabaleshwarkar, and Yoshi Suhara. 2025. When2Call: When (not) to Call Tools. In 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), pages 3391–3409, Albuquerque, New Mexico. Association for Computational Linguistics. https://aclanthology.org/2025.naacl-long.174/ (Call / refuse / clarify / answer decision.)
3. In-Vehicle Task-Oriented Dialogue — Huifang Du, Shuqin Li, Yi Dai, and Haofen Wang. 2025. Effortless In-Vehicle Task-Oriented Dialogue: Enhancing Natural and Efficient Interactions. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (CHI EA '25), Article 213, 1–10. ACM. https://doi.org/10.1145/3706599.3719799 (Multi-turn in-vehicle dialogue structure.)
4. CAR-bench — Johannes Kirmayr, Lukas Stappen, and Elisabeth André. 2026. CAR-bench: Evaluating the Consistency and Limit-Awareness of LLM Agents under Real-World Uncertainty. arXiv:2601.22027. https://arxiv.org/abs/2601.22027 (Hallucination / limit-awareness and disambiguation tasks — inspiration for the negative subsets.)
5. CTFusion — Daniel Rim, Minsoo Cho, Changwoo Chun, and Jaegul Choo. 2025. To Chat or Task: a Multi-turn Dialogue Generation Framework for Task-Oriented Dialogue Systems. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track), pages 576–592, Vienna, Austria. Association for Computational Linguistics. https://aclanthology.org/2025.acl-industry.41/ (Multi-turn chat/task dialogue generation and persona/slot augmentation.)

Citation

@misc{nova_iva_toolcalls_2026,
  title        = {Nova IVA Tool-Calls v0.2: A Synthetic Dataset for
                  In-Vehicle Assistant Tool Calling},
  author       = {Senthil Kumar N},
  year         = {2026},
  howpublished = {https://huggingface.co/datasets/Senthi1Kumar/nova-iva-toolcalls-v0.2},
  note         = {Synthetic English tool-calling dataset with schema-validated
                  function calls, when2call behaviors, and a held-out
                  novel-tool split for zero-shot generalization.}
}

License

Released for research use. The dataset is fully synthetic (no scraped user data). Generated with third-party LLMs via OpenRouter; downstream users are responsible for compliance with those providers' terms. Not affiliated with, or endorsed by, any vehicle manufacturer or model provider.