File size: 9,618 Bytes
27c779e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
# AppWorld-Vox

Synthetic English audio for AppWorld tasks, with hesitation, correction and interruption variants. Each sample contains the task setup, task annotation, and **at most one selected WAV per type**. Start with `manifest.jsonl` to load audio; use [the format reference](docs/FORMAT.md) to interpret annotations and timing.

This export has **189 samples from 63 families and 1,275 WAVs**. Train contains 54 samples / 371 WAVs; test contains 135 samples / 904 WAVs. Split assignments inherit the existing task split, with no family shared between train and test. Clean baseline audio is included for all 189 samples, using the same per-sample speaker as the existing variants.

## Available audio

| Type | WAVs | Expected control intent | Definition |
|---|---:|---|---|
| clean | 189 | execute | Original normalized instruction with no generated fillers or inserted pauses. |
| filler | 181 | execute | Hesitations/fillers or added pauses preserve the original request. |
| false-start | 189 | execute | A mistaken value is corrected immediately within the same utterance. |
| addition | 174 | execute | An initially omitted condition or argument is added after the initial request. |
| revision | 179 | execute | A value in a completed initial request is corrected in a later utterance. |
| cancel | 182 | cancel | A later utterance cancels the entire request; rollback is not requested. |
| halt | 181 | wait | A later utterance asks the agent to pause and await further instructions. |

162 samples have all six perturbation types plus clean (seven conditions). The remaining 48 sample/type combinations are skipped because no eligible LLM-selected WAV was available; samples are retained when at least one type is available. `reports/missing.json` records every skip. These are not 1,275 independent tasks: several audio variants share one sample's task and database.

Audio is 24 kHz, mono, 16-bit PCM WAV. All audio was copied unchanged from the selected generation output. Other types can also include fillers: labels describe the primary perturbation. See [statistics.json](statistics.json) for duration, speaker-ID and split statistics.

## Directory layout

```text
Appworld-vox/
├── README.md
├── manifest.jsonl                 # One record per WAV; canonical index
├── splits.json                    # Sample IDs in each split
├── statistics.json                # Counts and audio properties
├── docs/FORMAT.md                 # Fields, timing coordinates, provenance
├── tools/validate_dataset.py      # Standalone standard-library validator
├── reports/                       # Selection, skips, hashes, verification
├── train/
│   └── <task_id>/
│       ├── specs.json             # Authoritative instruction and task setup
│       ├── dbs/                   # Task database snapshots
│       ├── ground_truth/          # Evaluation code and answers
│       ├── annotation.json        # Pause/correction candidates and spoken text
│       ├── variants.json          # Available types and missing_types
│       ├── audio/<type>.wav       # Up to seven WAVs; missing types have no file
│       └── metadata/<type>.json   # TTS, filler/pause events and LLM selection
└── test/
    └── <task_id>/                 # Same layout as train
```

All manifest paths are relative to this root. Absolute paths inside provenance metadata refer to the original generation environment and are not required to read this export.

## Read one sample

Run this from the dataset root using Python 3.9 or newer; no third-party package is needed:

```python
import json
import wave
from pathlib import Path

root = Path.cwd()
with (root / "manifest.jsonl").open() as f:
    row = next(json.loads(line) for line in f if line.strip())
metadata = json.loads((root / row["metadata"]).read_text())
print(row["task_id"], row["type"], row["expected_control_intent"])
print(metadata["source_metadata"]["spoken_utterances"])
with wave.open(str(root / row["audio"]), "rb") as wav:
    pcm_bytes = wav.readframes(wav.getnframes())
    sample_rate = wav.getframerate()
```

Filter manifest rows by `split`, `type` or `task_id` as needed. For a comparison across all six conditions, group by `task_id` and retain only IDs with all six perturbation types (162 samples); these also have clean audio. Do not fill missing conditions with another sample's audio.

## Selection and task sources

For the six perturbation types, selection follows the LLM default recommendation implemented by the annotation-review backend at export time. Within each sample/type, choose the best naturalness rank that passes current-source and WAV-hash checks and Audio semantic equivalence. False-start, addition and revision additionally require Text semantic equivalence. Rank 1 is skipped if it fails these gates; singleton candidates receive the same semantic checks. Unknown/uncertain judgments are not accepted. Human Yes/No reviews are preserved but do not drive this LLM selection. Clean originals were subsequently imported at user request from `tasks_llm/<task_id>/audio_natural/clean/`; each has exactly one source candidate. Their generation ASR and quality flags are preserved, but this import did not add LLM semantic or human listening approval. Clean metadata uses `llm_selection: null` and records `import_selection` instead. No new LLM or TTS generation was run during export.

Task assets normally come from the original `tasks` directory; annotations and audio come from `tasks_llm`. Six tasks use existing TTS-aligned instruction overrides from `tasks_tts_1`, and three of those also use synchronized evaluation answers. [reports/task_source_overrides.json](reports/task_source_overrides.json) preserves the exact before/after documents and hashes. Use each sample's `specs.json` as the instruction authority.

Clean and the first four perturbation types target task execution with the original final goal. Cancel and halt require separate control-behavior evaluation: passing or failing the original completion test alone does not establish correct cancellation or waiting. `ground_truth/` is evaluation material, not agent input. Running tasks requires a compatible AppWorld runtime, which is not bundled here.

## Filler positions and metadata

`metadata/<type>.json` preserves the actual TTS utterances, source annotation/variant documents, ASR/WER, speaker and generation settings, LLM judgments and ranking reasons. `source_metadata.utterance_reports[].timeline` retains filler labels, neighboring words, boundary context and alignment flags. `inserted_pause_intervals` gives added short/long silence spans in **final WAV frames and seconds**, with exclusive end positions.

Inserted silence spans are not filler speech spans. ASR word times use source-utterance coordinates, while inserted intervals use final-WAV coordinates. Exact filler word boundaries are not human verified. See [docs/FORMAT.md](docs/FORMAT.md) before using timing for alignment or playback overlays.

## Verify the export

From this directory, run:

```bash
python3 tools/validate_dataset.py
```

Alternatively, pass a relocated dataset root as the first argument. The tool checks manifest paths, WAV bytes/headers, metadata consistency, pause bounds, missing-type records, split membership and inventory hashes. It loads no model, follows no historical source paths and writes no files. A successful run prints JSON with `status: passed`; failures exit with a nonzero status.

`reports/validation.json` records original export verification. `reports/documentation_audit.json` records the subsequent standalone structural audit. Inventory hash exclusions are explicit in `reports/inventory_policy.json`. Verification does not certify human listening quality or task execution success.

The original dataset was replaced only after verification, and its temporary backup was removed as requested; `reports/lifecycle.json` records this. `reports/rebuild_appworld_vox.py` is a historical script requiring the original internal sources, not a portable generation pipeline. Embedded source documents remain available if those original paths are unavailable. No new license is assigned by this export; license and citation information were not supplied.

Clean import: [reports/clean_import.json](reports/clean_import.json) lists all 189 sources, destination records and validation results. Earlier selection, lifecycle and documentation audit reports describe the initial six-type export; the latest sample-update report records the current seven-condition validation. Existing perturbation WAVs, including previously identified semantic issues, were not replaced or reselected.

Sample audio refresh (2026-10-19): `5a83b05_2` and `9871968_3` were updated from the regenerated source outputs, including clean. The six perturbation types use the latest eligible annotation-review LLM defaults; clean uses the regenerated original. Annotation normalization, metadata, pause intervals and manifest hashes were refreshed. See [reports/sample_update_20261019.json](reports/sample_update_20261019.json) for before/after hashes, actual utterances and current validation. Earlier snapshot/import reports are historical.

Ranked repair refresh: selected sample/type audio and metadata were updated after current text/audio semantic review and old/new naturalness comparison. See [reports/ranked_repair_20261020T020225128139Z.json](reports/ranked_repair_20261020T020225128139Z.json) for the selected sources and previous/new WAV hashes.