Lucasllfs/router-assistant-tool-calling-en-es

Router Assistant Tool Calling EN-ES

Synthetic English and Spanish conversations for supervised fine-tuning of a small, local router assistant. The assistant answers brief social turns, obtains current network facts through tools, handles tool failures, and asks for confirmation before restarting the router or disabling WAN internet access.

Dataset size

Split	Conversations	Assistant completions
Train	11,066	21,242
Validation	984	1,890
Test	926	1,769
Total	12,976	24,901

The corpus contains 7,006 English and 5,970 Spanish conversations. All rows from one synthetic generation batch are assigned to the same split using provenance.generation_id; there is no record-ID or generation-group overlap.

Configurations

• huggingface: conversational messages + tools. Tool-result contents are canonical JSON strings, while function arguments remain JSON objects for native chat templates.
• openai_wire: API transport representation; function arguments are canonical JSON strings.
• trl_completion: one row per assistant turn, with prompt, completion, and production tool schemas. Use completion-only loss and disable packing.

from datasets import load_dataset

conversations = load_dataset(
    "Lucasllfs/router-assistant-tool-calling-en-es",
    "huggingface",
)

trl_rows = load_dataset(
    "Lucasllfs/router-assistant-tool-calling-en-es",
    "trl_completion",
)

The original canonical JSONL is retained at source/router_assistant_v1.jsonl for provenance and future exports. It is not part of a Viewer configuration because its tool-result content field intentionally contains structured JSON objects while ordinary message content is text.

Tools

The router_core_v1 registry defines:

• get_internet_status()
• run_speed_test()
• get_connected_device_count()
• restart_router()
• set_internet_enabled(enabled)

The complete JSON Schemas are embedded in every training row and published at metadata/tool_registry.json.

Generation and validation

User utterances were generated synthetically with stealth/ox-alpha. A deterministic local renderer produced tool calls, tool results, errors, confirmation trajectories, and grounded final answers. The promoted build contains 17,628 attempted examples: 12,976 were accepted, 296 failed semantic validation, and 4,356 exact or near duplicates were removed.

Automated checks reported:

• 12,976/12,976 canonical records valid;
• zero exact and near user-text duplicates in the promoted corpus;
• 100% valid tool names, arguments, and call/result linkage;
• all required quality gates passed;
• 55/55 pipeline tests passed before publication.

See reports/quality-report.md, docs/DATASET.md, and the manifests under metadata/ for details and SHA-256 hashes.

Intended use

The dataset is intended for narrow-domain SFT and evaluation of compact assistants that operate a router through a host-controlled tool layer. Candidate models should be evaluated with their native chat templates and on the target Qualcomm device.

The host application must enforce authorization and confirmation. The model is not a security boundary. In particular, the host must reject router restart or WAN-disable calls unless it holds a valid confirmation state for that exact action.

Limitations

• The corpus is synthetic and does not replace a human-written gold evaluation set.
• English and Spanish should receive native-speaker review before production release.
• Router APIs, result schemas, and wording may require adaptation to the actual device.
• Performance on other networking tasks, languages, or tool registries is untested.
• The dataset's license: other metadata is intentionally conservative. Review the applicable OpenRouter and teacher-model terms before redistributing or changing the repository to public visibility.