Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
metric_source: string
thresholds: struct<minimum_turn_duration_seconds: double, short_turn_max_asr_chunks: int64, negative_latency: st (... 5 chars omitted)
child 0, minimum_turn_duration_seconds: double
child 1, short_turn_max_asr_chunks: int64
child 2, negative_latency: string
metrics: struct<pause_handling: struct<samples: int64, take_over_count: int64, take_over_rate: double, mean_l (... 591 chars omitted)
child 0, pause_handling: struct<samples: int64, take_over_count: int64, take_over_rate: double, mean_latency_seconds_given_ta (... 126 chars omitted)
child 0, samples: int64
child 1, take_over_count: int64
child 2, take_over_rate: double
child 3, mean_latency_seconds_given_take_over: double
child 4, records: list<item: struct<sample_id: string, takes_turn: bool, latency: double>>
child 0, item: struct<sample_id: string, takes_turn: bool, latency: double>
child 0, sample_id: string
child 1, takes_turn: bool
child 2, latency: double
child 5, correct_wait_rate: double
child 1, smooth_turn_taking: struct<samples: int64, take_over_count: int64, take_over_rate: double, mean_latency_seconds_given_ta (... 99 chars omitted)
child 0, samples: int64
child 1, take_over_count: int64
child 2, take_over_rate: double
child 3, mean_latency_seconds_given_take_over: double
child 4, records: list<item: struct<sample_id: string, takes_turn: bool, latency: double>>
child 0, item: struct<sample_id: string, takes_turn: bool, latency: double>
child 0, sample_id: string
child 1, takes_turn: bool
child 2, latency: double
child 2, user_interruption: struct<samples: int64, take_over_count: int64, take_over_rate: double, mean_latency_seconds_given_ta (... 99 chars omitted)
child 0, samples: int64
child 1, take_over_count: int64
child 2, take_over_rate: double
child 3, mean_latency_seconds_given_take_over: double
child 4, records: list<item: struct<sample_id: string, takes_turn: bool, latency: double>>
child 0, item: struct<sample_id: string, takes_turn: bool, latency: double>
child 0, sample_id: string
child 1, takes_turn: bool
child 2, latency: double
not_transferred: struct<backchannel_jsd: string, user_interruption_relevance: string, v1_5_behavior: string>
child 0, backchannel_jsd: string
child 1, user_interruption_relevance: string
child 2, v1_5_behavior: string
text_without_word_chunks: int64
metric_evaluator: string
multiword_chunks: int64
event_outputs: int64
clean_outputs: int64
note: string
invalid_timestamp_files: int64
speech_boundaries: string
word_alignment: string
to
{'event_outputs': Value('int64'), 'clean_outputs': Value('int64'), 'word_alignment': Value('string'), 'speech_boundaries': Value('string'), 'invalid_timestamp_files': Value('int64'), 'text_without_word_chunks': Value('int64'), 'multiword_chunks': Value('int64'), 'metric_evaluator': Value('string'), 'note': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 489, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2818, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
metric_source: string
thresholds: struct<minimum_turn_duration_seconds: double, short_turn_max_asr_chunks: int64, negative_latency: st (... 5 chars omitted)
child 0, minimum_turn_duration_seconds: double
child 1, short_turn_max_asr_chunks: int64
child 2, negative_latency: string
metrics: struct<pause_handling: struct<samples: int64, take_over_count: int64, take_over_rate: double, mean_l (... 591 chars omitted)
child 0, pause_handling: struct<samples: int64, take_over_count: int64, take_over_rate: double, mean_latency_seconds_given_ta (... 126 chars omitted)
child 0, samples: int64
child 1, take_over_count: int64
child 2, take_over_rate: double
child 3, mean_latency_seconds_given_take_over: double
child 4, records: list<item: struct<sample_id: string, takes_turn: bool, latency: double>>
child 0, item: struct<sample_id: string, takes_turn: bool, latency: double>
child 0, sample_id: string
child 1, takes_turn: bool
child 2, latency: double
child 5, correct_wait_rate: double
child 1, smooth_turn_taking: struct<samples: int64, take_over_count: int64, take_over_rate: double, mean_latency_seconds_given_ta (... 99 chars omitted)
child 0, samples: int64
child 1, take_over_count: int64
child 2, take_over_rate: double
child 3, mean_latency_seconds_given_take_over: double
child 4, records: list<item: struct<sample_id: string, takes_turn: bool, latency: double>>
child 0, item: struct<sample_id: string, takes_turn: bool, latency: double>
child 0, sample_id: string
child 1, takes_turn: bool
child 2, latency: double
child 2, user_interruption: struct<samples: int64, take_over_count: int64, take_over_rate: double, mean_latency_seconds_given_ta (... 99 chars omitted)
child 0, samples: int64
child 1, take_over_count: int64
child 2, take_over_rate: double
child 3, mean_latency_seconds_given_take_over: double
child 4, records: list<item: struct<sample_id: string, takes_turn: bool, latency: double>>
child 0, item: struct<sample_id: string, takes_turn: bool, latency: double>
child 0, sample_id: string
child 1, takes_turn: bool
child 2, latency: double
not_transferred: struct<backchannel_jsd: string, user_interruption_relevance: string, v1_5_behavior: string>
child 0, backchannel_jsd: string
child 1, user_interruption_relevance: string
child 2, v1_5_behavior: string
text_without_word_chunks: int64
metric_evaluator: string
multiword_chunks: int64
event_outputs: int64
clean_outputs: int64
note: string
invalid_timestamp_files: int64
speech_boundaries: string
word_alignment: string
to
{'event_outputs': Value('int64'), 'clean_outputs': Value('int64'), 'word_alignment': Value('string'), 'speech_boundaries': Value('string'), 'invalid_timestamp_files': Value('int64'), 'text_without_word_chunks': Value('int64'), 'multiword_chunks': Value('int64'), 'metric_evaluator': Value('string'), 'note': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
GPT-Realtime on Vi-FDB v1: full reference outputs
Public reference outputs from GPT-Realtime on all 400 Vi-FDB event cases and their 200 available clean controls. This repository contains model outputs and evaluation evidence, not benchmark inputs.
- Canonical benchmark: https://huggingface.co/datasets/tuanamz/vi-fdb-v1
- Benchmark revision:
c9730eddf78b094610be09796afe8e60c178caef - Interactive explorer: https://huggingface.co/spaces/tuanamz/vi-fdb-v1-explorer
Contents
- 400 event-condition GPT-Realtime output WAVs.
- 200 clean-control GPT-Realtime output WAVs.
- PhoWhisper-large transcripts with word-level timestamps for the full expansion and faithful timing-metric artifacts for all 400 event outputs.
- Silero VAD speech boundaries used by the original FDB timing metrics.
- Sanitized shared-clock timing events.
- Automated GPT-4.1-mini behavior verdicts.
- 29 explicit native-speaker adjudications.
- Aggregate semantic evidence and faithfully reproduced original FDB v1.0 metrics.
Raw transport logs, API credentials, machine-local paths, vendor session IDs, vendor response IDs, and copies of benchmark input audio are deliberately excluded. Each row links to its canonical benchmark input at the frozen revision above.
Semantic result: pilot, expansion, and full benchmark
| Task | Pilot | Expansion | Full |
|---|---|---|---|
| Backchannel | 100% | 100% | 100% |
| Pause handling | 40% | 36.7% | 38% |
| Smooth turn-taking | 85% | 73.3% | 78% |
| User interruption — standard | 40% | 80% | 64% |
| Background speech | 30% | 10% | 18% |
| Talking to other | 30% | 26.7% | 28% |
| User backchannel | 100% | 96.7% | 98% |
| User interruption — paired variant | 80% | 96.7% | 90% |
| Overall | 63.1% (101/160) | 65.0% (156/240) | 64.2% (257/400) |
The pilot includes 29 native-speaker corrections. The expansion uses automated role-aware GPT-4.1-mini verdicts and still requires an equivalent human audit.
Original FDB v1.0 metrics — all 50 cases per task
| Metric | Result |
|---|---|
| Pause take-over rate (lower is better) | 48% (24/50) |
| Pause correct-wait rate | 52% (26/50) |
| Smooth-turn take-over rate | 98% (49/50) |
| Smooth-turn latency, conditional on response | 1.000 s |
| Post-interruption response rate | 76% (38/50) |
| Post-interruption latency, conditional on response | 0.662 s |
These numbers were reproduced with the original Full-Duplex-Bench v1.0 formulas and thresholds. PhoWhisper-large supplies Vietnamese lexical units; Silero VAD supplies observable speech onset and offset because raw Whisper alignment can stretch the first word across leading silence. The literal upstream pause and smooth-turn scripts reproduce the reported rates and latency. The original English backchannel JSD is not reported because its ICC reference distribution is language/corpus specific.
Row schema
data/metadata.jsonl contains one row per generated output condition. The
file_name column points to generated audio in data/audio/. Important fields
include:
- benchmark repository, frozen revision, task, sample ID, and source suite;
- event or clean condition;
- model name and PhoWhisper-large transcript;
- word-level ASR chunks, VAD timing provenance, and sanitized timing events as JSON strings;
- automated and final pass/fail values for event-condition rows;
- automated or human adjudication source;
- observed behavior, judge confidence, and Vietnamese judge evidence;
- direct links to canonical benchmark input and metadata.
Clean-control rows intentionally have no pass/fail verdict because the published role-aware verdict applies to the event-condition case.
Interpretation and reuse
Use this repository to audit one reference submission, test Vi-FDB tooling, or compare a new model on the same frozen benchmark revision. Do not treat the generated audio or transcripts as benchmark ground truth. Do not train on these outputs and then report an unqualified score on the same pilot.
The package follows the canonical benchmark's CC BY-NC 4.0 license. The audio is model-generated; applicable model/service terms may impose additional conditions.
- Downloads last month
- 51