README.md

---
license: mit
task_categories:
  - audio-classification
tags:
  - audio
  - tool-calling
  - function-calling
  - benchmark
pretty_name: Audio2Tool Private Eval Set
extra_gated_prompt: >-
  This is the held-out evaluation set for the Audio2Tool Leaderboard.
  By requesting access you agree to use it only for evaluating models and
  submitting to the leaderboard, not to redistribute the audio, and not to
  attempt to reconstruct or publish the hidden ground-truth labels.
extra_gated_fields:
  Affiliation: text
  Intended use: text
---

# Audio2Tool Private Eval Set

Held-out evaluation set for the
[Audio2Tool Leaderboard](https://huggingface.co/spaces/RVtech/Audio2ToolLeaderboard).
It is a random subset of the private split of the
[Audio2Tool dataset](https://huggingface.co/datasets/RVtech/Audio2Tool):
up to 200 queries per tier across all 8 tiers (1,579 samples).
**Ground-truth labels are not published** — submissions are scored on the
leaderboard Space against hidden labels.

## Contents

- `metadata.jsonl` — one row per sample: `sample_id`, `tier`, `audio`
  (list of wav paths, ordered turns for `tier7_multiturn`)
- `audio/<tier>/query_XXXXX/*.wav` — audio files

Tools available to the model are described in
[`tools_registry.csv`](https://huggingface.co/datasets/RVtech/Audio2Tool/blob/main/tools_registry.csv)
of the main dataset.

## How to submit

1. For each row in `metadata.jsonl`, run your model on the audio and produce
   a tool-call prediction string, e.g.
   `setZoneTemperature(zone=Driver, temperature=21.0)`.
   For multi-intent samples, output multiple calls in one string, first call
   is the primary tool:
   `setLockState(state=Locked) setFanSpeed(level=7)`
2. Write predictions as JSONL, one row per sample:

```json
{"sample_id": "tier1_direct/00042", "prediction": "setZoneTemperature(zone=Driver, temperature=21.0)"}
```

3. Upload the file on the **Submit** tab of the
   [leaderboard Space](https://huggingface.co/spaces/RVtech/Audio2ToolLeaderboard).

## Evaluation

For each sample the leaderboard computes (same normalization as the
[benchmark code](https://github.com/audio2tool/Audio2Tool)):

- **Tool Accuracy** — predicted primary tool name matches ground truth
  (case-insensitive)
- **Exact Match** — tool and all parameters match exactly
- **Param F1** — F1 over predicted vs ground-truth parameter key/value pairs
  (0 when the tool is wrong)

Scores are reported per tier plus a macro-average across tiers.

## Running with the benchmark code

The [Audio2Tool benchmark repo](https://github.com/audio2tool/Audio2Tool)
can produce predictions for the supported models; the eval set follows the
same layout as the public release, so the `release` dataset loader reads
`metadata.jsonl` directly.