Yambda-5B -- A Large-Scale Multi-modal Dataset for Ranking And Retrieval
Paper • 2505.22238 • Published • 4
Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed because of a cast error
Error code: DatasetGenerationCastError
Exception: DatasetGenerationCastError
Message: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 7 new columns ({'NDCG@10', 'NDCG@10_lo95', 'NDCG@10_hi95', 'NDCG@20_hi95', 'NDCG@20_lo95', 'method', 'NDCG@20'}) and 7 missing columns ({'delta_lo95', 'K', 'id_method', 'audio_method', 'delta_mean', 'p_one_sided', 'delta_hi95'}).
This happened while the csv dataset builder was generating data using
hf://datasets/Vladislavbro-500/music-recommendations/eval_results/summary.csv (at revision 6c164325241e552f3f9f7a9f591e55221d3bae07), ['hf://datasets/Vladislavbro-500/music-recommendations@6c164325241e552f3f9f7a9f591e55221d3bae07/eval_results/paired.csv', 'hf://datasets/Vladislavbro-500/music-recommendations@6c164325241e552f3f9f7a9f591e55221d3bae07/eval_results/summary.csv', 'hf://datasets/Vladislavbro-500/music-recommendations@6c164325241e552f3f9f7a9f591e55221d3bae07/eval_results/summary_by_size.csv']
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1848, in _prepare_split_single
writer.write_table(table)
~~~~~~~~~~~~~~~~~~^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 765, in write_table
self._write_table(pa_table, writer_batch_size=writer_batch_size)
~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 773, in _write_table
pa_table = table_cast(pa_table, self._schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
method: string
NDCG@10: double
NDCG@10_lo95: double
NDCG@10_hi95: double
NDCG@20: double
NDCG@20_lo95: double
NDCG@20_hi95: double
-- schema metadata --
pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 1119
to
{'audio_method': Value('string'), 'id_method': Value('string'), 'K': Value('int64'), 'delta_mean': Value('float64'), 'delta_lo95': Value('float64'), 'delta_hi95': Value('float64'), 'p_one_sided': Value('float64')}
because column names don't match
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1850, in _prepare_split_single
raise DatasetGenerationCastError.from_cast_error(
...<4 lines>...
)
datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 7 new columns ({'NDCG@10', 'NDCG@10_lo95', 'NDCG@10_hi95', 'NDCG@20_hi95', 'NDCG@20_lo95', 'method', 'NDCG@20'}) and 7 missing columns ({'delta_lo95', 'K', 'id_method', 'audio_method', 'delta_mean', 'p_one_sided', 'delta_hi95'}).
This happened while the csv dataset builder was generating data using
hf://datasets/Vladislavbro-500/music-recommendations/eval_results/summary.csv (at revision 6c164325241e552f3f9f7a9f591e55221d3bae07), ['hf://datasets/Vladislavbro-500/music-recommendations@6c164325241e552f3f9f7a9f591e55221d3bae07/eval_results/paired.csv', 'hf://datasets/Vladislavbro-500/music-recommendations@6c164325241e552f3f9f7a9f591e55221d3bae07/eval_results/summary.csv', 'hf://datasets/Vladislavbro-500/music-recommendations@6c164325241e552f3f9f7a9f591e55221d3bae07/eval_results/summary_by_size.csv']
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
audio_method string | id_method string | K int64 | delta_mean float64 | delta_lo95 float64 | delta_hi95 float64 | p_one_sided float64 |
|---|---|---|---|---|---|---|
AudioAGREE | AGREE | 10 | 0.011043 | 0.005281 | 0.017677 | 0 |
AudioAGREE | AGREE | 20 | 0.011245 | 0.005249 | 0.01763 | 0 |
AudioAGREE | GroupIM | 10 | 0.008991 | 0.002378 | 0.015438 | 0.005 |
AudioAGREE | GroupIM | 20 | 0.010136 | 0.003743 | 0.016401 | 0.001 |
GroupCrossAttn | AGREE | 10 | 0.012541 | 0.00541 | 0.019391 | 0 |
GroupCrossAttn | AGREE | 20 | 0.012334 | 0.005946 | 0.018964 | 0 |
GroupCrossAttn | GroupIM | 10 | 0.010489 | 0.00329 | 0.017521 | 0.001 |
GroupCrossAttn | GroupIM | 20 | 0.011225 | 0.004617 | 0.017814 | 0 |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
null | null | null | null | null | null | null |
Group aggregators over a frozen scorer — artifacts (YAMBDA-50m)
Чекпоинты и промежуточные артефакты для групповых музыкальных рекомендаций:
per-user скорер, кэш его топ-200, аудиоэмбеддинги каталога и обученные
групповые агрегаторы. Всё производное от
yandex/yambda, flavor 50m.
Состав
| Путь | Что это |
|---|---|
gsasrec/best.pt, config.json, metrics.csv |
SASRec-скорер, 276 305 items, test NDCG@10 = 0.0726 |
gsasrec/item_id_to_idx.pkl |
item_id YAMBDA → компактный индекс |
user_scores_cache/scores.parquet |
топ-200 скоров на пользователя, 9 170 пользователей |
audio/embeddings.npy |
аудиоэмбеддинги YAMBDA, 276 306 × 128, переиндексированные под каталог |
audio/user_profiles.npy, uid_to_row.pkl, user_audio_valid.npy |
аудиопрофили пользователей и маска покрытия |
aggregators/{agree,groupim,audio_agree,group_cross_attn}/ |
чекпоинты групповых агрегаторов |
aggregators/groups_split.pkl |
синтетический split групп (10k/2k/2k, размеры 2–5, seed 42) |
eval_results/ |
summary.csv, summary_by_size.csv, paired.csv, per_sample.npz, LaTeX-таблица |
Результаты (групповой NDCG на test-группах)
| Метод | NDCG@10 | NDCG@20 |
|---|---|---|
| GroupCrossAttn | 0.0916 | 0.1024 |
| Audio-AGREE | 0.0901 | 0.1014 |
| Max Pleasure | 0.0828 | 0.0915 |
| GroupIM | 0.0811 | 0.0912 |
| AGREE | 0.0790 | 0.0901 |
| Average | 0.0695 | 0.0819 |
| Least Misery | 0.0312 | 0.0363 |
Скорер заморожен, обучаются только параметры агрегатора. Кандидаты берутся из кэша топ-200, поэтому числа сравнимы между методами, но усечены по построению.
Лицензия и атрибуция
Apache 2.0, унаследована от исходного датасета. audio/embeddings.npy —
отфильтрованная и переиндексированная копия аудиоэмбеддингов YAMBDA
(Copyright 2025 YANDEX LLC). YAMBDA опубликован для научных и исследовательских
целей; эти производные распространяются на тех же условиях.
Препринт датасета: https://arxiv.org/abs/2505.22238
- Downloads last month
- 128