Datasets:

ankit-pn
/

phone-assistant-function-calling

license: cc-by-4.0
language:
  - en
  - hi
task_categories:
  - text-generation
tags:
  - function-calling
  - tool-use
  - phone-assistant
  - hinglish
  - benchmark
pretty_name: Phone Assistant Function Calling (20 tools, EN + Hinglish)
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
      - split: test
        path: test.jsonl

Phone Assistant Function Calling

This benchmark has 4,000 voice-assistant style commands, each labelled with calls to 20 phone tools (tools.json). It was built to fine-tune and evaluate FunctionGemma 270M.

Composition: 3,000 single calls (150 per function), 480 parallel requests with 2–3 independent calls, and 520 no-call requests (small talk and unsupported actions).
Languages: 85% English and 15% Hinglish (Hindi in Latin script mixed with English).
Split: 3,200 train / 800 test, stratified. Groups of near-duplicate utterances never cross the split.
Authoring: every utterance and label was written individually by LLM writers following SPEC.md. No sentence templates or slot filling were used. Separate LLM reviewers audited the records, 32 bad labels were excluded, and normalized duplicates were removed.

Record format

{"id": "...", "text": "kal subah 7 baje ka alarm laga do", "calls": [{"name": "set_alarm", "arguments": {"time": "07:00"}}], "lang": "hi-en", "category": "single", "group": "..."}
{"id": "...", "text": "thanks, you're the best", "calls": [], "lang": "en", "reply": "You're welcome!", "category": "no_call", "group": "..."}

Argument conventions: 24h HH:MM times, dates as lower-case phrases, durations in seconds or minutes, and optional arguments only when the user states them. The full rules are in SPEC.md.

Baselines (test split)

Model	Exact match	Tool selection	No-call recall
google/functiongemma-270m-it	22.3%	31.6%	90.4%
ankit-pn/functiongemma-270m-phone-assistant	75.4%	93.0%	79.8%

Code and scorer: https://github.com/ankit-pn/functiongemma-phone-assistant-bench

Limitations: this is synthetic, LLM-authored data and not real user logs. Exact match undercounts acceptable paraphrases in free-text fields.