conversations string | question_audios list | answer_audios list |
|---|---|---|
[{"role": "user", "content": "Give three tips for staying healthy."}, {"role": "assistant", "content": "Sure! Here are three tips for staying healthy: 1. Eat a balanced diet rich in fruits, vegetables, lean proteins, and whole grains. This helps ensure your body gets the necessary nutrients. 2. Exercise regularly, ai... | [] | [
[
1049,
811,
1626,
1652,
1335,
347,
1234,
152,
61,
188,
500,
451,
1145,
1727,
637,
1911,
1420,
828,
915,
1710,
1576,
1454,
1610,
1401,
299,
860,
294,
577,
1360,
406,
1075,
1748,
731,
... |
[{"role": "user", "content": "Describe the structure of an atom."}, {"role": "assistant", "content": "An atom is composed of three main particles: protons, neutrons, and electrons. Protons and neutrons are located in the nucleus at the center of the atom. The nucleus is very small but contains most of the atom's mass. ... | [] | [
[
1049,
316,
1626,
1620,
267,
347,
825,
1744,
2035,
1785,
121,
245,
66,
815,
1434,
1078,
337,
1408,
1979,
1339,
987,
212,
293,
56,
683,
1986,
1510,
1068,
731,
1411,
277,
1571,
1508,
... |
[{"role": "user", "content": "How can we reduce air pollution?"}, {"role": "assistant", "content": "Reducing air pollution requires a combination of individual actions and systemic changes. We can start by using public transportation, biking, or walking instead of driving cars to reduce vehicle emissions. Supporting an... | [] | [
[
1049,
477,
1626,
142,
537,
11,
712,
754,
958,
1549,
1369,
643,
1652,
766,
533,
1017,
1388,
1032,
1707,
1100,
766,
709,
378,
94,
170,
1753,
1894,
109,
1004,
1708,
318,
1098,
54,
1... |
[{"role": "user", "content": "Describe a time when you had to make a difficult decision."}, {"role": "assistant", "content": "There was a time when I had to choose between moving to a new city for a great job opportunity and staying in my hometown to support my family. It was a challenging decision because the job offe... | [] | [
[
1049,
477,
457,
1562,
1067,
555,
1978,
1648,
958,
1646,
919,
1289,
537,
1055,
47,
1017,
1746,
637,
411,
153,
383,
50,
1434,
1462,
12,
277,
429,
500,
953,
2029,
68,
698,
1001,
113... |
[{"role": "user", "content": "Write a short story in third person narration about a protagonist who has to make an important career decision."}, {"role": "assistant", "content": "Jane had always been passionate about art, but her parents had insisted on a more practical path. With a degree in business, she landed a sta... | [] | [
[
1049,
811,
1626,
546,
1335,
165,
1978,
1868,
958,
1965,
522,
1936,
1541,
487,
1247,
25,
1053,
1215,
121,
2031,
533,
818,
1082,
29,
921,
1715,
694,
1516,
1795,
271,
1223,
63,
957,
... |
[{"role": "user", "content": "Render a 3D model of a house"}, {"role": "assistant", "content": "I am a voice assistant and cannot write code. To render a 3D model of a house, first, you need to decide on the software or programming environment you want to use, such as Blender or a 3D graphics library. Then, create a ba... | [] | [
[
1049,
477,
1456,
390,
1067,
2010,
1044,
1744,
1272,
10,
1603,
1260,
272,
1080,
697,
95,
1494,
705,
1793,
2007,
1588,
916,
1362,
444,
1070,
1984,
1645,
516,
1322,
888,
958,
400,
778,
... |
[{"role": "user", "content": "How did Julius Caesar die?"}, {"role": "assistant", "content": "Julius Caesar was assassinated. He was stabbed to death by a group of Roman senators on the Ides of March, which is March 15, in 44 BC. The assassination took place at a meeting of the Senate at the Theatre of Pompey. Key figu... | [] | [
[
1049,
316,
1626,
142,
267,
11,
1044,
1744,
958,
2010,
1027,
2006,
716,
1727,
1504,
391,
795,
1190,
1483,
979,
154,
531,
1809,
1617,
1063,
1875,
1098,
176,
976,
671,
515,
677,
1825,
... |
[{"role": "user", "content": "What is the capital of France?"}, {"role": "assistant", "content": "The capital of France is Paris."}] | [] | [
[
1049,
1700,
1626,
142,
306,
555,
666,
1648,
958,
1438,
742,
1164,
273,
1519,
845,
156,
2045,
1418,
1219,
416,
1981,
438,
188,
710,
1662,
1248,
1673,
1776,
1041,
214,
1686,
1445,
1899... |
[{"role": "user", "content": "Generate a list of ten items a person might need for a camping trip"}, {"role": "assistant", "content": "Sure! Here are ten items someone might need for a camping trip: 1. Tent 2. Sleeping bag 3. Camping stove 4. Headlamp or flashlight 5. First aid kit 6. Water bottle or hydration system... | [] | [
[
1049,
811,
1626,
1562,
306,
1273,
1433,
152,
1876,
188,
1301,
151,
1696,
165,
731,
181,
783,
1816,
2036,
619,
1937,
242,
1730,
1442,
1184,
1735,
153,
910,
67,
406,
414,
1530,
731,
... |
[{"role": "user", "content": "Discuss the causes of the Great Depression"}, {"role": "assistant", "content": "The Great Depression was caused by a combination of factors. The stock market crash of 1929 was a significant trigger, causing a loss of confidence and financial turmoil. This led to a reduction in consumer spe... | [] | [
[
1049,
477,
1626,
142,
1335,
347,
825,
1190,
958,
1777,
323,
1057,
1652,
1940,
1237,
1321,
809,
220,
1679,
331,
1695,
1026,
1134,
1520,
1363,
786,
1738,
1002,
1521,
236,
209,
1803,
16... |
[{"role": "user", "content": "Explain the use of word embeddings in Natural Language Processing"}, {"role": "assistant", "content": "Word embeddings are a technique used in Natural Language Processing to represent words as vectors in a continuous vector space. This approach allows words with similar meanings to have si... | [] | [
[
1049,
811,
1626,
546,
1335,
347,
43,
1511,
1415,
1564,
1115,
977,
439,
1403,
1137,
1281,
168,
257,
239,
1019,
48,
1896,
1187,
926,
1373,
277,
209,
1921,
1312,
211,
1550,
1913,
1593,
... |
[{"role": "user", "content": "Describe the function of a computer motherboard"}, {"role": "assistant", "content": "The motherboard is the main circuit board in a computer. It provides the essential connections for all the components, including the CPU, memory, storage devices, and expansion cards, to communicate with e... | [] | [
[
1049,
477,
373,
2010,
267,
555,
1433,
1665,
958,
1519,
1121,
1289,
279,
1273,
384,
1551,
419,
690,
1916,
377,
419,
830,
1041,
1923,
1686,
354,
695,
1091,
494,
764,
227,
939,
1407,
... |
[{"role": "user", "content": "Propose an ethical solution to the problem of data privacy"}, {"role": "assistant", "content": "An ethical solution to the problem of data privacy can start with transparency. Organizations should clearly communicate how they collect, use, and store personal data. They should obtain explic... | [] | [
[
1049,
477,
1456,
390,
1067,
2010,
1044,
1744,
2035,
1216,
360,
215,
1274,
36,
1019,
1519,
337,
690,
1858,
165,
437,
750,
211,
1998,
125,
1376,
1361,
1590,
731,
207,
365,
77,
1508,
... |
[{"role": "user", "content": "Generate three verbs that mean the same as \"to apologize\""}, {"role": "assistant", "content": "Three verbs that mean the same as \"to apologize\" are \"to confess,\" \"to admit,\" and \"to express regret.\""}] | [] | [
[
1049,
1700,
373,
1562,
915,
1350,
1238,
754,
738,
1862,
355,
2027,
2005,
1455,
1234,
582,
357,
1701,
1007,
647,
1815,
645,
773,
1109,
357,
1984,
936,
105,
315,
2003,
991,
1103,
107,
... |
[{"role": "user", "content": "What does DNA stand for?"}, {"role": "assistant", "content": "DNA stands for Deoxyribonucleic Acid."}] | [] | [
[
1049,
477,
1626,
142,
1335,
555,
825,
1744,
1578,
743,
400,
1057,
1332,
1244,
1504,
776,
1207,
1984,
441,
1184,
849,
974,
1351,
402,
1661,
1215,
1201,
377,
1911,
1380,
508,
26,
1070,... |
[{"role": "user", "content": "Compare and contrast the Cuban Missile Crisis and the Vietnam War."}, {"role": "assistant", "content": "The Cuban Missile Crisis and the Vietnam War were both significant events during the Cold War, but they were different in their nature and outcomes. The Cuban Missile Crisis in 1962 was ... | [] | [
[
1049,
887,
374,
1868,
420,
11,
1238,
1180,
958,
893,
337,
1529,
490,
1506,
1039,
13,
1746,
345,
1679,
551,
969,
62,
1071,
1700,
28,
1611,
68,
917,
1163,
1007,
240,
1545,
785,
181... |
[{"role": "user", "content": "Generate a list of random words."}, {"role": "assistant", "content": "I am a voice assistant and cannot write code. To generate a list of random words, you can: 1. Make a list of words you'd like to choose from. 2. Use an online random word generator to select words for you. 3. If you hav... | [] | [
[
1049,
316,
373,
1620,
1067,
1443,
845,
1511,
1272,
344,
712,
177,
655,
1001,
547,
120,
1494,
1984,
1834,
1184,
350,
723,
542,
521,
1070,
1875,
403,
472,
1187,
319,
1691,
130,
778,
... |
[{"role": "user", "content": "Who is the world's most famous painter?"}, {"role": "assistant", "content": "Many consider Leonardo da Vinci to be the world's most famous painter, especially known for masterpieces like the Mona Lisa and The Last Supper."}] | [] | [
[
1049,
477,
1626,
142,
537,
11,
1978,
754,
958,
1336,
1208,
193,
1933,
731,
19,
1848,
1777,
172,
1679,
1418,
987,
1966,
497,
1139,
998,
277,
939,
742,
987,
1684,
1658,
2023,
998,
... |
[{"role": "user", "content": "Explain the concept of cogging torque."}, {"role": "assistant", "content": "Cogging torque is an undesirable effect found in electric motors, particularly permanent magnet motors. It occurs due to the interaction between the permanent magnets on the rotor and the stator slots' edges. This ... | [] | [
[
1049,
316,
1626,
142,
1335,
347,
43,
1868,
391,
1264,
505,
353,
1311,
859,
1388,
1500,
1023,
266,
268,
1994,
128,
267,
1834,
1612,
173,
2033,
1858,
1900,
822,
1647,
1381,
117,
1353,
... |
[{"role": "user", "content": "Look up the boiling point of water."}, {"role": "assistant", "content": "The boiling point of water is 100 degrees Celsius, which is equivalent to 212 degrees Fahrenheit, at standard atmospheric pressure."}] | [] | [
[
1049,
811,
1626,
290,
306,
1572,
712,
424,
958,
14,
457,
1477,
1335,
1567,
1039,
156,
419,
1319,
1603,
1134,
24,
1812,
1262,
1400,
9,
1168,
1149,
350,
557,
677,
1567,
977,
680,
1... |
[{"role": "user", "content": "Describe the life and reign of King Charles II."}, {"role": "assistant", "content": "King Charles II of England, Scotland, and Ireland was born on May 29, 1630. He became king following the tumultuous period of the English Civil War and the execution of his father, King Charles I. After sp... | [] | [
[
1049,
316,
1626,
1477,
1532,
332,
1433,
1511,
958,
1511,
1150,
1389,
599,
1176,
1989,
1576,
1207,
1984,
1135,
1741,
112,
465,
1588,
362,
194,
1450,
17,
656,
1862,
1526,
858,
1220,
22... |
[{"role": "user", "content": "Identify the lines of longitude that divides North and South America."}, {"role": "assistant", "content": "The line of longitude that divides North and South America is generally considered to be the 60th meridian west. This line runs north-south and is used as a reference point. Additiona... | [] | [
[
1049,
887,
374,
1868,
1773,
714,
1752,
1808,
958,
1646,
325,
97,
1298,
347,
513,
1868,
419,
888,
1098,
783,
879,
293,
1553,
946,
383,
1701,
1957,
1171,
646,
136,
810,
703,
1136,
... |
TinyOmni SFT Data
Training data for TinyOmni, a small multimodal model (text / audio / image β text / audio). Three Parquet files covering the text-to-speech, image-to-text, and speech conversation stages of the seven-step SFT pipeline. Audio is stored not as waveforms but as 8-layer residual vector quantization (RVQ) codebook indices.
| Config | File | Rows | Size | Stage |
|---|---|---|---|---|
sft_text_to_audio |
sft_text_to_audio.parquet |
1,248,923 | 4.99 GB | Text instruction β speech reply |
sft_image_to_text |
sft_image_to_text.parquet |
2,904,511 | 4.60 GB | Image + text β text reply |
sft_audio_to_audio |
sft_audio_to_audio.parquet |
414,024 | 5.36 GB | Speech conversation β speech reply |
Repository structure
The three data files live flat at the repository root (no subdirectories). Each config maps to a single Parquet file.
FLCgeigeigei/tinyomni-data/
βββ .gitattributes
βββ LICENSE
βββ README.md
βββ sft_text_to_audio.parquet
βββ sft_image_to_text.parquet
βββ sft_audio_to_audio.parquet
Dataset structure
Each sample is one row. A sample combines a text conversation with optional audio tokens and image bytes that are aligned to the conversation turns:
sample
ββ conversations JSON list of messages (user / assistant alternating)
ββ question_audios input audio, one entry per user turn (empty in text-to-audio)
ββ answer_audios reply audio, one entry per assistant turn
β ββ each entry = interleaved 8-layer RVQ token sequence (flat list<int64>)
ββ ref_audios reference audio tokens (interleaved 8-layer), for speaker cloning
ββ spk_emb 192-dim speaker embedding
ββ image_bytes raw JPEG (image-to-text config only)
Turn alignment is the key structural rule: the k-th element of answer_audios
corresponds to the k-th assistant message in conversations, and the k-th element
of question_audios corresponds to the k-th user message. Training supervises only
the last assistant turn.
sft_text_to_audio β text instruction β speech reply
| Field | Type | Shape (one sample) |
|---|---|---|
conversations |
string | 2 messages: 1 user + 1 assistant |
question_audios |
list<binary> | empty list |
answer_audios |
list<list<int64>> | 1 entry Γ 2400 tokens (= 8 layers Γ 300 frames) |
// conversations
[
{"role": "user", "content": "Give three tips for staying healthy."},
{"role": "assistant", "content": "Sure! Here are three tips ..."}
]
// answer_audios: [[token_0, token_1, ..., token_2399]]
// -> 1 assistant turn, one 2400-token interleaved sequence
sft_image_to_text β image + text β text reply
| Field | Type | Shape (one sample) |
|---|---|---|
conversations |
large_string | 2 messages: 1 user + 1 assistant |
image_bytes |
large_binary | raw JPEG, ~5 KB |
[
{"role": "user", "content": "<image>How does the use of negative space and color contrast contribute to the overall impact?"},
{"role": "assistant", "content": "The use of negative space ..."}
]
The <image> placeholder position varies between samples (before or after the question
text), so the loader must treat it as an in-text marker rather than a fixed prefix.
This config has no audio fields.
sft_audio_to_audio β speech conversation β speech reply
| Field | Type | Shape (one sample) |
|---|---|---|
conversations |
string | multi-turn user / assistant messages |
question_audios |
list<binary> | 1 entry per user turn |
answer_audios |
list<list<int64>> | 1 entry per assistant turn, e.g. 1536 tokens (= 8 Γ 192 frames) |
ref_audios |
list<int64> | 1344 tokens (= 8 Γ 168 frames), interleaved |
spk_emb |
list<double> | 192 dims |
// conversations
[
{"role": "user", "content": "How can I interact with you to get optimum performance?"},
{"role": "assistant", "content": "Hello! You can interact with me by ..."}
]
// question_audios: [<audio_bytes>] -> 1 user turn, 1 input audio
// answer_audios: [[token_0, ..., token_1535]] -> 1 assistant turn, 1536 tokens
// ref_audios: [token_0, ..., token_1343] -> reference voice, interleaved 8-layer
// spk_emb: [0.047, 0.985, ...] -> 192 floats
Schema
| Column | Type | Aligned to | Description |
|---|---|---|---|
conversations |
string / large_string | β | JSON-encoded message list (user / assistant) |
question_audios |
list<binary> | user turns | Input audio bytes; empty in sft_text_to_audio |
answer_audios |
list<list<int64>> | assistant turns | Reply audio; each entry is an interleaved 8-layer RVQ token sequence |
image_bytes |
large_binary | β | Raw JPEG bytes (sft_image_to_text only) |
ref_audios |
list<int64> | β | Reference audio tokens, interleaved 8-layer |
spk_emb |
list<double> | β | Speaker embedding, 192 dims |
Audio representation
Audio is stored as 8-layer residual vector quantization (RVQ) codebook indices, not waveforms. Each audio field is a single flat integer sequence where the 8 layers are interleaved per timestep:
[t0_L0, t0_L1, ..., t0_L7, t1_L0, t1_L1, ..., t1_L7, ...]
so a sequence of length N corresponds to N / 8 audio frames across 8 codebook layers.
- Vocabulary size: 2051
2050is the end-of-sequence (stop) token-100is an ignore label used on the training side and generally does not appear in the data files
Usage
from datasets import load_dataset
ds = load_dataset(
"FLCgeigeigei/tinyomni-data",
data_files="sft_image_to_text.parquet",
split="train",
)
print(ds.features)
For per-column reading with PyArrow (recommended for the large binary/audio columns):
import pyarrow.parquet as pq
t = pq.read_table("sft_text_to_audio.parquet", columns=["conversations"])
License
Released under CC BY 4.0.
Please credit FLCgeigeigei/tinyomni-data when using this data.
- Downloads last month
- 85