DUSUNEN Model Lab — Open Turkish AI Models & Live Demos
Collection
Open Turkish AI by Göktuğ Düşünen: OCR, speech, text restoration, embeddings, reranking, benchmarks and live demos. • 18 items • Updated
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
model: string
parameters: int64
embedding_dimension: int64
TurHistQuadRetrieval: double
XQuADRetrieval: double
WebFAQRetrieval: double
MKQARetrieval: double
BelebeleRetrieval: double
macro_average: double
tasks: list<item: struct<name: string, dataset: struct<path: string, revision: string>, license: string, do (... 89 chars omitted)
child 0, item: struct<name: string, dataset: struct<path: string, revision: string>, license: string, domains: list (... 77 chars omitted)
child 0, name: string
child 1, dataset: struct<path: string, revision: string>
child 0, path: string
child 1, revision: string
child 2, license: string
child 3, domains: list<item: string>
child 0, item: string
child 4, eval_splits: list<item: string>
child 0, item: string
child 5, subsets: list<item: string>
child 0, item: string
raw_mteb_result: struct<model_name: string, model_revision: string, task_results: list<item: struct<dataset_revision: (... 13073 chars omitted)
child 0, model_name: string
child 1, model_revision: string
child 2, task_results: list<item: struct<dataset_revision: string, task_name: string, mteb_version: string, scores: struct< (... 12954 chars omitted)
child 0, item: struct<dataset_revision: string, task_name: string, mteb_version: string, scores: struct<test: list< (... 12942 chars omitted)
child 0, dataset_revision: string
child 1, task_name: string
child
...
double
child 147, hit_rate_at_1000: double
child 148, main_score: double
child 149, hf_subset: string
child 150, languages: list<item: string>
child 0, item: string
child 151, mteb_version: string
child 4, evaluation_time: double
child 5, kg_co2_emissions: null
child 6, date: string
child 7, evaluation_phases: list<item: struct<name: string, start: double, end: double, split: string, subset: string>>
child 0, item: struct<name: string, start: double, end: double, split: string, subset: string>
child 0, name: string
child 1, start: double
child 2, end: double
child 3, split: string
child 4, subset: string
child 3, exceptions: list<item: null>
child 0, item: null
child 4, experiment_name: null
device: string
model_revision: null
language_filter: string
prompt_style: string
mteb_version: string
normalized_embeddings: bool
inference_dtype: string
model_source: string
suite: string
task_main_scores: struct<TurHistQuadRetrieval: double, XQuADRetrieval: double, WebFAQRetrieval: double, MKQARetrieval: (... 35 chars omitted)
child 0, TurHistQuadRetrieval: double
child 1, XQuADRetrieval: double
child 2, WebFAQRetrieval: double
child 3, MKQARetrieval: double
child 4, BelebeleRetrieval: double
to
{'suite': Value('string'), 'language_filter': Value('string'), 'model': Value('string'), 'model_source': Value('string'), 'model_revision': Value('null'), 'parameters': Value('int64'), 'embedding_dimension': Value('int64'), 'inference_dtype': Value('string'), 'normalized_embeddings': Value('bool'), 'prompt_style': Value('string'), 'device': Value('string'), 'mteb_version': Value('string'), 'task_main_scores': {'TurHistQuadRetrieval': Value('float64'), 'XQuADRetrieval': Value('float64'), 'WebFAQRetrieval': Value('float64'), 'MKQARetrieval': Value('float64'), 'BelebeleRetrieval': Value('float64')}, 'macro_average': Value('float64'), 'tasks': List({'name': Value('string'), 'dataset': {'path': Value('string'), 'revision': Value('string')}, 'license': Value('string'), 'domains': List(Value('string')), 'eval_splits': List(Value('string')), 'subsets': List(Value('string'))}), 'raw_mteb_result': {'model_name': Value('string'), 'model_revision': Value('string'), 'task_results': List({'dataset_revision': Value('string'), 'task_name': Value('string'), 'mteb_version': Value('string'), 'scores': {'test': List({'ndcg_at_1': Value('float64'), 'ndcg_at_3': Value('float64'), 'ndcg_at_5': Value('float64'), 'ndcg_at_10': Value('float64'), 'ndcg_at_20': Value('float64'), 'ndcg_at_100': Value('float64'), 'ndcg_at_1000': Value('float64'), 'map_at_1': Value('float64'), 'map_at_3': Value('float64'), 'map_at_5': Value('float64'), 'map_at_10': Value('float64'), 'map_at_20': Value('float64'), 'map_at_1
...
float64'), 'nauc_mrr_at_1_diff1': Value('float64'), 'nauc_mrr_at_3_max': Value('float64'), 'nauc_mrr_at_3_std': Value('float64'), 'nauc_mrr_at_3_diff1': Value('float64'), 'nauc_mrr_at_5_max': Value('float64'), 'nauc_mrr_at_5_std': Value('float64'), 'nauc_mrr_at_5_diff1': Value('float64'), 'nauc_mrr_at_10_max': Value('float64'), 'nauc_mrr_at_10_std': Value('float64'), 'nauc_mrr_at_10_diff1': Value('float64'), 'nauc_mrr_at_20_max': Value('float64'), 'nauc_mrr_at_20_std': Value('float64'), 'nauc_mrr_at_20_diff1': Value('float64'), 'nauc_mrr_at_100_max': Value('float64'), 'nauc_mrr_at_100_std': Value('float64'), 'nauc_mrr_at_100_diff1': Value('float64'), 'nauc_mrr_at_1000_max': Value('float64'), 'nauc_mrr_at_1000_std': Value('float64'), 'nauc_mrr_at_1000_diff1': Value('float64'), 'hit_rate_at_1': Value('float64'), 'hit_rate_at_3': Value('float64'), 'hit_rate_at_5': Value('float64'), 'hit_rate_at_10': Value('float64'), 'hit_rate_at_20': Value('float64'), 'hit_rate_at_100': Value('float64'), 'hit_rate_at_1000': Value('float64'), 'main_score': Value('float64'), 'hf_subset': Value('string'), 'languages': List(Value('string')), 'mteb_version': Value('string')})}, 'evaluation_time': Value('float64'), 'kg_co2_emissions': Value('null'), 'date': Value('string'), 'evaluation_phases': List({'name': Value('string'), 'start': Value('float64'), 'end': Value('float64'), 'split': Value('string'), 'subset': Value('string')})}), 'exceptions': List(Value('null')), 'experiment_name': Value('null')}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, 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 127, 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 483, 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 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 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
model: string
parameters: int64
embedding_dimension: int64
TurHistQuadRetrieval: double
XQuADRetrieval: double
WebFAQRetrieval: double
MKQARetrieval: double
BelebeleRetrieval: double
macro_average: double
tasks: list<item: struct<name: string, dataset: struct<path: string, revision: string>, license: string, do (... 89 chars omitted)
child 0, item: struct<name: string, dataset: struct<path: string, revision: string>, license: string, domains: list (... 77 chars omitted)
child 0, name: string
child 1, dataset: struct<path: string, revision: string>
child 0, path: string
child 1, revision: string
child 2, license: string
child 3, domains: list<item: string>
child 0, item: string
child 4, eval_splits: list<item: string>
child 0, item: string
child 5, subsets: list<item: string>
child 0, item: string
raw_mteb_result: struct<model_name: string, model_revision: string, task_results: list<item: struct<dataset_revision: (... 13073 chars omitted)
child 0, model_name: string
child 1, model_revision: string
child 2, task_results: list<item: struct<dataset_revision: string, task_name: string, mteb_version: string, scores: struct< (... 12954 chars omitted)
child 0, item: struct<dataset_revision: string, task_name: string, mteb_version: string, scores: struct<test: list< (... 12942 chars omitted)
child 0, dataset_revision: string
child 1, task_name: string
child
...
double
child 147, hit_rate_at_1000: double
child 148, main_score: double
child 149, hf_subset: string
child 150, languages: list<item: string>
child 0, item: string
child 151, mteb_version: string
child 4, evaluation_time: double
child 5, kg_co2_emissions: null
child 6, date: string
child 7, evaluation_phases: list<item: struct<name: string, start: double, end: double, split: string, subset: string>>
child 0, item: struct<name: string, start: double, end: double, split: string, subset: string>
child 0, name: string
child 1, start: double
child 2, end: double
child 3, split: string
child 4, subset: string
child 3, exceptions: list<item: null>
child 0, item: null
child 4, experiment_name: null
device: string
model_revision: null
language_filter: string
prompt_style: string
mteb_version: string
normalized_embeddings: bool
inference_dtype: string
model_source: string
suite: string
task_main_scores: struct<TurHistQuadRetrieval: double, XQuADRetrieval: double, WebFAQRetrieval: double, MKQARetrieval: (... 35 chars omitted)
child 0, TurHistQuadRetrieval: double
child 1, XQuADRetrieval: double
child 2, WebFAQRetrieval: double
child 3, MKQARetrieval: double
child 4, BelebeleRetrieval: double
to
{'suite': Value('string'), 'language_filter': Value('string'), 'model': Value('string'), 'model_source': Value('string'), 'model_revision': Value('null'), 'parameters': Value('int64'), 'embedding_dimension': Value('int64'), 'inference_dtype': Value('string'), 'normalized_embeddings': Value('bool'), 'prompt_style': Value('string'), 'device': Value('string'), 'mteb_version': Value('string'), 'task_main_scores': {'TurHistQuadRetrieval': Value('float64'), 'XQuADRetrieval': Value('float64'), 'WebFAQRetrieval': Value('float64'), 'MKQARetrieval': Value('float64'), 'BelebeleRetrieval': Value('float64')}, 'macro_average': Value('float64'), 'tasks': List({'name': Value('string'), 'dataset': {'path': Value('string'), 'revision': Value('string')}, 'license': Value('string'), 'domains': List(Value('string')), 'eval_splits': List(Value('string')), 'subsets': List(Value('string'))}), 'raw_mteb_result': {'model_name': Value('string'), 'model_revision': Value('string'), 'task_results': List({'dataset_revision': Value('string'), 'task_name': Value('string'), 'mteb_version': Value('string'), 'scores': {'test': List({'ndcg_at_1': Value('float64'), 'ndcg_at_3': Value('float64'), 'ndcg_at_5': Value('float64'), 'ndcg_at_10': Value('float64'), 'ndcg_at_20': Value('float64'), 'ndcg_at_100': Value('float64'), 'ndcg_at_1000': Value('float64'), 'map_at_1': Value('float64'), 'map_at_3': Value('float64'), 'map_at_5': Value('float64'), 'map_at_10': Value('float64'), 'map_at_20': Value('float64'), 'map_at_1
...
float64'), 'nauc_mrr_at_1_diff1': Value('float64'), 'nauc_mrr_at_3_max': Value('float64'), 'nauc_mrr_at_3_std': Value('float64'), 'nauc_mrr_at_3_diff1': Value('float64'), 'nauc_mrr_at_5_max': Value('float64'), 'nauc_mrr_at_5_std': Value('float64'), 'nauc_mrr_at_5_diff1': Value('float64'), 'nauc_mrr_at_10_max': Value('float64'), 'nauc_mrr_at_10_std': Value('float64'), 'nauc_mrr_at_10_diff1': Value('float64'), 'nauc_mrr_at_20_max': Value('float64'), 'nauc_mrr_at_20_std': Value('float64'), 'nauc_mrr_at_20_diff1': Value('float64'), 'nauc_mrr_at_100_max': Value('float64'), 'nauc_mrr_at_100_std': Value('float64'), 'nauc_mrr_at_100_diff1': Value('float64'), 'nauc_mrr_at_1000_max': Value('float64'), 'nauc_mrr_at_1000_std': Value('float64'), 'nauc_mrr_at_1000_diff1': Value('float64'), 'hit_rate_at_1': Value('float64'), 'hit_rate_at_3': Value('float64'), 'hit_rate_at_5': Value('float64'), 'hit_rate_at_10': Value('float64'), 'hit_rate_at_20': Value('float64'), 'hit_rate_at_100': Value('float64'), 'hit_rate_at_1000': Value('float64'), 'main_score': Value('float64'), 'hf_subset': Value('string'), 'languages': List(Value('string')), 'mteb_version': Value('string')})}, 'evaluation_time': Value('float64'), 'kg_co2_emissions': Value('null'), 'date': Value('string'), 'evaluation_phases': List({'name': Value('string'), 'start': Value('float64'), 'end': Value('float64'), 'split': Value('string'), 'subset': Value('string')})}), 'exceptions': List(Value('null')), 'experiment_name': Value('null')}}
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.
DUSUNEN Turkish Retrieval Benchmark v1
A reproducible five-task Turkish retrieval evaluation package. It references the original public datasets instead of republishing their content and stores pinned revisions, exact language subsets, raw MTEB result objects and a compact comparison table.
Why five tasks?
The original DUSUNEN evaluation used only Ottoman-history questions. This suite broadens the evidence across reading comprehension, web FAQ, open-domain knowledge questions and multilingual comprehension while preserving a fully Turkish evaluation path.
| Task | Turkish subset | Domain | Main metric |
|---|---|---|---|
| TurHistQuad Retrieval | default |
history / encyclopaedic QA | nDCG@10 |
| XQuAD Retrieval | tr |
reading comprehension | nDCG@10 |
| WebFAQ Retrieval | tur |
web FAQ | nDCG@10 |
| MKQA Retrieval | tr |
open-domain knowledge QA | nDCG@10 |
| Belebele Retrieval | tur_Latn-tur_Latn |
reading comprehension | nDCG@10 |
Protocol
mteb==2.18.16- Pinned dataset revisions in
suite-manifest.json - Turkish-only exclusive subsets
- Normalized embeddings
- Explicit per-model prompt format (
plain,harrierore5) - Shared official MTEB retrieval evaluator
- BF16 inference on the same RTX 5060 Laptop GPU
- No task text is used for training or hard-negative mining
Raw per-task metric dictionaries are kept in each model result JSON. The comparison table is generated from those files, not typed by hand.
Published result summary
| Rank | Model | Parameters | Embedding dim | Five-task macro |
|---|---|---|---|---|
| 1 | multilingual E5 base | 278,043,648 | 768 | 0.619618 |
| 2 | DUSUNEN Rota 270M v2 | 268,098,176 | 640 | 0.568602 |
| 3 | DUSUNEN Rota 270M v1 | 268,098,176 | 640 | 0.565896 |
| 4 | DUSUNEN Pusula 118M v0 | 117,653,760 | 384 | 0.480070 |
The hard-negative v2 continuation improved v1 on all five tasks, but its macro
gain was only 0.002706. E5 remains the overall suite leader, so this release
does not make a state-of-the-art claim. The complete analysis, including the
frozen-candidate reranking experiment, is available in
report/README.md and as a print-ready HTML document in
report/index.html.
Limitations
- The tasks are primarily question-to-passage retrieval and do not cover every Turkish search intent.
- The macro average weights tasks equally regardless of corpus size.
- Some datasets are translations, so translation quality can affect results.
- A benchmark suite is evidence about these tasks, not a universal quality guarantee.
- Downloads last month
- 65