Dataset Viewer
Auto-converted to Parquet Duplicate
question_id
stringlengths
29
29
history_id
stringlengths
22
26
root_sample_id
stringlengths
13
13
leakage_group_id
stringlengths
19
19
taxonomy
stringclasses
4 values
question
stringlengths
26
156
answer
stringlengths
24
137
distance_bin
stringclasses
3 values
query_time_seconds
float64
1.34k
17.8k
question_audio
stringlengths
86
86
answer_audio
stringlengths
84
84
history
listlengths
8
109
evidence_json
stringlengths
2
6.45k
history_audio_seconds
float64
957
2.28k
facteval_8a20783e8f098d5e1b58
POD0000000231_W000_T00
POD0000000231
LG_475cf5cb19b7f48f
multi-session-fact
What was your first novel called, and which of your books did Ant-Man read on screen?
My first novel was Looking for Alaska; Ant-Man read The Fault in Our Stars.
near
2,645.46
audio/question/d6/d67177e43f8f6ebf1ba742d170a8aa980a8807b0177f88e3c0e2ab41ff5ad882.wav
audio/answer/e8/e87b919a72391936a2708dd2155a8f7eee2b0a37879500fb185a7d3672193c3c.wav
[ { "assistant": "I usually say that I work in educational video so that they don't ask more questions. But if I'm feeling a little more talkative, I I usually say that I'm a a writer and a YouTuber. Right. But I guess now I'm also a podcaster, so I don't know what I do.", "assistant_audio": "audio/history/ac...
[{"turn_index": 43, "role": "assistant", "speaker": "SPEAKER_01", "session_id": "POD0000000231_W000_T00_S004", "source_turn_indices": [54], "start_seconds": 2341.74, "end_seconds": 2367.96, "query_gap_seconds": 277.5, "position_bin": "near", "quote": "When it's things like Ant-Man reading The Fault in Our Stars, it's j...
1,473.69
facteval_ef24ffebd645b8a493e1
POD0000000406_W000_T00
POD0000000406
LG_dbf67b8ef0050afb
multi-session-fact
What did I call your childhood play dates, and what did I say you and your future husband were discussing by the end of your first evening?
You called them anthropological experiments and said we ended that first evening discussing when we would get married.
near
1,562.553
audio/question/5c/5c00951c84bced8c5d6e7bf4b1ba01c5ffd725f391df36e97902747adb7794b3.wav
audio/answer/0f/0f8574ae1d60863a3f792e87ec94877c25ea74d5b895a7e4ee7529a5b357f734.wav
[ { "assistant": "Well, I would start with the fact that uh The day I was born, my mother received a telegram from her husband who had sailed to the UK to enlist, the war having begun in Europe and not yet the United States wasn't yet involved. And he sent off a cable saying, \"Congratulations on baby Catherine. ...
[{"turn_index": 50, "role": "user", "speaker": "SPEAKER_00", "session_id": "POD0000000406_W000_T00_S003", "source_turn_indices": [72], "start_seconds": 1357.1, "end_seconds": 1409.31, "query_gap_seconds": 153.243, "position_bin": "near", "quote": "Hmm. And one thing that's so interesting in the vision of life that you'...
1,233.093
facteval_755bc0d2951d09612b68
POD0000000406_W000_T01
POD0000000406
LG_dbf67b8ef0050afb
absence
In which Iranian city did I say you spent time before the revolution?
The Iranian city was not identified in our conversation.
null
2,995.32
audio/question/95/9511bcf287bd5b22ba95608ea09a3a5edeac3ff1de5b6e305c42b619b6040a1d.wav
audio/answer/79/79e0a7f1000f2e77d08af3a7d945c07100df84652b514b4e322468ac56595cbc.wav
[ { "assistant": "Well, you've probably seen that the the slogan that gets quoted all the time, never doubt that a small group of thoughtful committed citizens can change the world.", "assistant_audio": "audio/history/e5/e5a5dccde1f0d33c12950eea8ea03effa6cc0112a40bf49de1360961af77b73d.wav", "overlap_type"...
[]
1,239.15
facteval_f9b7d54bb44d08cc7e96
POD0000000417_W000_T00
POD0000000417
LG_dbf67b8ef0050afb
single-user-fact
Which book did I connect with the discipline of living and reflecting?
You connected that discipline with Composing a Life.
far
1,591.1
audio/question/5b/5bca069e94172231d71564de447831e26adae6fac398b31fe905bc1f8494e1b5.wav
audio/answer/94/941f0880323e6cbe445b4bdd037688e450bab81990f4000974445683b2685843.wav
[ { "assistant": "Well, I would start with the fact that uh The day I was born, my mother received a telegram from her husband who had sailed to the UK to enlist, the war having begun in Europe and not yet, the United States wasn't yet involved. And he sent off a cable saying, \"Congratulations on baby Catherine....
[{"turn_index": 12, "role": "user", "speaker": "SPEAKER_00", "session_id": "POD0000000417_W000_T00_S000", "source_turn_indices": [24], "start_seconds": 517.47, "end_seconds": 551.2, "query_gap_seconds": 1039.9, "position_bin": "far", "quote": "And I think your book Composing a Life I mean, that is a phrase that you bro...
1,260.16
facteval_e5a9490e411dc2148b31
POD0000000417_W000_T01
POD0000000417
LG_dbf67b8ef0050afb
single-user-fact
What animal did I recall your father taking to his opening classes at the San Francisco Art Institute?
You recalled my father taking a crab to his opening classes.
middle
2,995.22
audio/question/36/36698e76403b3b0c6cfaa67cc7c9974fd6c1d8ff5c8529c6e74d7f6dd301fa63.wav
audio/answer/66/66fc8727510bb14931e0821d9779c5262332c70e0b07e4f43128ca549d237670.wav
[ { "assistant": "Well, you've probably seen the the the slogan that gets quoted all the time. Never doubt that a small group of thoughtful, committed citizens can change the world.", "assistant_audio": "audio/history/0f/0ff13aa8ce63eec7d536643796ab5612e3634014c242ec505c3dc5cfb99ffc5a.wav", "overlap_type"...
[{"turn_index": 32, "role": "user", "speaker": "SPEAKER_00", "session_id": "POD0000000417_W000_T01_S002", "source_turn_indices": [58], "start_seconds": 2470.81, "end_seconds": 2523.2, "query_gap_seconds": 472.02, "position_bin": "middle", "quote": "You wrote a beautiful beautiful essay about um your father's death in 1...
1,238.42
facteval_afddb93e9f989b5226db
POD0000000420_W000_T00
POD0000000420
LG_51805737df13577c
single-user-fact
Which of your poetry collections about the end of your marriage did I say won the Pulitzer Prize?
You said my collection Stag's Leap won the Pulitzer Prize.
near
1,544.38
audio/question/2a/2ac62a01f15938593c9f2286d79a3799f5578e5ed13104654287114241b8214b.wav
audio/answer/34/34e1351a45a0d40ba26800543418888349597281b85b8bdb8c25ffbc6d57e14c.wav
[ { "assistant": "Yes. Thank you. Thank you for having me. Um, well, I would say that even before I was born, I was a pagan dancer, I would say. I was upside down and I was I was engaged in rhythm. I don't remember this, but it just seems true that the heartbeat and the breath and the orborigmy, the sounds of thi...
[{"turn_index": 34, "role": "user", "speaker": "SPEAKER_01", "session_id": "POD0000000420_W000_T00_S003", "source_turn_indices": [54, 55], "start_seconds": 1153.18, "end_seconds": 1222.96, "query_gap_seconds": 321.42, "position_bin": "near", "quote": "And a lot of them are here. Poetry reading men. Um, okay. Um, it's s...
1,368.16
facteval_a06739b6e6817e6e0d6a
POD0000000420_W000_T01
POD0000000420
LG_51805737df13577c
single-assistant-fact
In the poem No Makeup, how did the speaker feel when the makeup came off?
She felt like a small mammal in the woods with a speaking countenance.
far
3,005.12
audio/question/c8/c8116dbb448d7c0b4279887d1406835722e76ae59c39688067492d8732ab7e94.wav
audio/answer/3f/3f3cc6f0e5dd882a1770975bfb6f6277420d121e17f70688ae170cb35114a4f4.wav
[{"assistant":"You","assistant_audio":"audio/history/0e/0e8c016f19fcd0b8e1dfdafce14ce582ca14f14028c8(...TRUNCATED)
"[{\"turn_index\": 7, \"role\": \"assistant\", \"speaker\": \"SPEAKER_02\", \"session_id\": \"POD000(...TRUNCATED)
1,229.24
facteval_f1b7ced4a706904b97a3
POD0000000424_W000_T00
POD0000000424
LG_d6fb933517b22363
multi-session-fact
"Which two New York organizations did I introduce you as leading, and which drag character did I lat(...TRUNCATED)
You named Lab Shul and the Storytelling Project, then asked about Hadassah Gross.
far
1,548.852
audio/question/c1/c115eea38000409d5334d632b1a83579ff80ef90c436e1d8c6f47210bc58d4ea.wav
audio/answer/05/05fe3a2ae509cba64d5cd3882f6632003b7129eff21d05ae9c666d875a2735a6.wav
[{"assistant":"Hmm. So my task on Fridays growing up in preparation for the Sabbath, um, was to help(...TRUNCATED)
"[{\"turn_index\": 0, \"role\": \"user\", \"speaker\": \"SPEAKER_01\", \"session_id\": \"POD00000004(...TRUNCATED)
1,265.422
facteval_2e0f0b46f52b0dbe1e5d
POD0000000424_W000_T01
POD0000000424
LG_d6fb933517b22363
absence
What was the name of the cemetery you visited for the second anniversary of your father's death?
The cemetery's name was not provided.
null
3,095.9
audio/question/62/62f47d8f2ceda016f259dd296238a959eb806a92d73fdc6d653b9567d0b13711.wav
audio/answer/ad/ad294515db4fd0a8b4e8cdfa592efeeb387f4a778b7ba8ef7bca15cb8b25333d.wav
[{"assistant":"That was the","assistant_audio":"audio/history/c4/c4f683d9b75dde150f4bd78230af7a01960(...TRUNCATED)
[]
1,383.5
facteval_3c126230fc0e160eea2a
POD0000000433_W000_T00
POD0000000433
LG_b2697ccabcf138a8
single-user-fact
"Which book did I cite for Reinhold Niebuhr's line about humanity being its own most vexing problem?(...TRUNCATED)
The line came from The Nature and Destiny of Man.
far
1,544.04
audio/question/24/24db09867be314995b1a641bb4eb83eabfc39602363591432900f78f3b785470.wav
audio/answer/da/dad59d6258a89ec9f104e439c0dac374885052bd11668d363f0aa69202af0828.wav
[{"assistant":"Well, I suppose I was blessed by being born into an amazing landscape in the west of (...TRUNCATED)
"[{\"turn_index\": 8, \"role\": \"user\", \"speaker\": \"SPEAKER_02\", \"session_id\": \"POD00000004(...TRUNCATED)
1,360.31
End of preview. Expand in Data Studio

GigaSpeech Eval500 — Spoken Conversation Memory QA

Custom held-out evaluation data used with shared-ttt-giga-fixed-multiqa: 500 histories and 500 questions, from release multiqa_fixed_r20261019. This is a project-specific memory-QA split, not the official GigaSpeech ASR test set.

Data

Question type Count
User fact 200
Assistant fact 150
Multi-session fact 100
Absence 50

There are 14,383 history exchanges. Both sides of each released history have their original WAV audio and transcript. User/assistant are replay roles assigned to conversation speakers. Question and reference-answer speech are synthesized. The package contains 29,728 unique WAV files, approximately 22.13 GB / 192.12 audio hours, all 16 kHz mono 16-bit PCM. These hours sum individual audio files and are not a wall-clock session-duration measurement.

Audio is stored unchanged in approximately 1 GiB TAR shards under audio_archives/. audio_manifest.jsonl maps every relative WAV path to an archive member, SHA256, duration and role. No resampling or lossy encoding was applied.

Download and use

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="Xinnian/gigaspeech-eval500",
    repo_type="dataset",
    local_dir="gigaspeech-eval500",
)
python gigaspeech-eval500/prepare_local.py --root gigaspeech-eval500

The script verifies and extracts all archive members, then writes local/eval.histories.json, local/eval.query.json and local/eval.stream.jsonl with absolute local paths and an updated history-catalog checksum. Source manifests in the repository stay unchanged. --verify-only checks archive members without extracting them or writing local manifests.

In the salmonn-omni-ttt evaluation configuration, point data.dataset_dir, evaluation.dataset_path and evaluation.metadata_path to the generated local/eval.query.json; point data.stream_manifest and evaluation.timeline_manifest to local/eval.stream.jsonl. Use absolute paths for these settings and retain data.timeline_input: streaming. The model and upstream dependency paths must also match your local installation.

The Parquet view can be read independently:

from datasets import load_dataset
data = load_dataset("Xinnian/gigaspeech-eval500", split="test")

Its audio columns contain package-relative path strings; the WAVs become accessible after downloading and extracting the archives. load_dataset by itself does not unpack the audio. Full text, history audio references, QA labels and evidence are in each row; real_overlap_metadata_json and evidence_json preserve nested annotations as JSON strings.

Evaluation layout

  • eval.histories.json: ordered paired exchanges, both audio paths, source indices and overlap flags.
  • eval.query.json: questions and gold answers for answer-only supervision/scoring.
  • eval.stream.jsonl: history + spoken current question only; gold answers are excluded from replay.
  • eval.qa.jsonl: full QA annotations, evidence, taxonomy and portable synthesis metadata.
  • data/eval.parquet: 500-row dataset view, exposed as the test split on the Hub.
  • dataset_manifest.json, audio_manifest.jsonl, SHA256SUMS: counts, source hashes, archive mapping and integrity information.

Reference-answer WAVs are provided for completeness and must not be included in the model's input history. Evidence turn indices and timestamps retain the original source annotation coordinate system; they are not necessarily direct indices into the paired old_turns list. The source allocator counters are preserved for provenance; recompute them if evaluating with a later allocator implementation.

Downloads last month
49