Pin dataset HF revisions for reproducibility (#37)

* Rename EUSIPCO paper

* Allow passing the HF revision for datasets loading

* uv lock

* set dataset revisions

* CHANGELOG

* Update braindecode

Author: PierreGtch

…_02_EUSIPCO_benchmark_FM_pro_license.pdf 2026_02_EUSIPCO_benchmark_FM.pdf2026_02_EUSIPCO_benchmark_FM_pro_license.pdf renamed to 2026_02_EUSIPCO_benchmark_FM.pdf 2026_02_EUSIPCO_benchmark_FM_pro_license.pdf renamed to 2026_02_EUSIPCO_benchmark_FM.pdf

@@ -8,9 +8,17 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## [Unreleased]
 
 ### Added
+- Add a required `revision` field on `Dataset` (forwarded to `BaseConcatDataset.pull_from_hub`) and pin every default dataset config to its current HuggingFace commit SHA for reproducibility ([#37](https://github.com/braindecode/OpenEEGBench/pull/37)).
 - Allow custom normalization methods by subclassing `Normalization`, now based on `exca.helpers.DiscriminatedModel`. Builtin subclasses pin their pre-existing `kind` value to preserve cached experiment UIDs ([#35](https://github.com/braindecode/OpenEEGBench/pull/35)).
 - Add `training_required_parameters` field on `_BackboneBase` for top-level `nn.Parameter` names that must remain trainable (e.g. Labram's `temporal_embedding`). Distinct from `training_required_modules`, which only accepts `nn.Module` names. Finetuning methods declare compatibility via the `supports_training_required_parameters` class variable; `IA3` and `OFT` opt out. `LoRA`/`AdaLoRA`/`DoRA` forward the list to PEFT's `target_parameters`; `Frozen`/`TwoStages` extend their unfreeze list ([#25](https://github.com/braindecode/OpenEEGBench/pull/25)).
 
+### Changed
+- X
+
+### Fixed
+- X
+
+
 ## [0.4.0] - 2026-05-07
 
 ### Added

CHANGELOG.md

@@ -101,6 +101,12 @@ class Dataset(BaseModel):
     hf_id: str = Field(
         description="HuggingFace Hub dataset ID, e.g. 'braindecode/bcic2a'."
     )
+    revision: str = Field(
+        description=(
+            "HuggingFace Hub revision (commit hash, branch, or tag) to pin "
+            "the dataset to a specific version."
+        ),
+    )
     n_classes: int | None = Field(
         description="Number of classes (None for regression)."
     )
@@ -113,15 +119,19 @@ class Dataset(BaseModel):
         ),
     )
     preload: bool = False
-    
+
     _exclude_from_cls_uid: ClassVar[tuple[str, ...]] = ("preload",)
 
     def load(self):
         """Pull windowed dataset from HuggingFace Hub."""
         from braindecode.datasets import BaseConcatDataset
 
-        log.info("Loading dataset from HuggingFace Hub: %s", self.hf_id)
-        return BaseConcatDataset.pull_from_hub(self.hf_id, preload=self.preload)
+        log.info(
+            "Loading dataset from HuggingFace Hub: %s@%s", self.hf_id, self.revision
+        )
+        return BaseConcatDataset.pull_from_hub(
+            self.hf_id, preload=self.preload, revision=self.revision
+        )
 
     def setup(self, normalization):
         """Load dataset, apply normalization, split, and return sets.

open_eeg_bench/dataset.py

@@ -16,6 +16,7 @@ def arithmetic_zyma2019(**overrides) -> Dataset:
     """Mental arithmetic vs. rest — 2 classes, 36 subjects."""
     defaults = dict(
         hf_id="braindecode/arithmetic_zyma2019",
+        revision="8359206e7e1b3d1f068033348e291d6677be1920",
         n_classes=2,
         montage_name="standard_1005",
         splitter=PredefinedSplitter(
@@ -33,6 +34,7 @@ def bcic2a(**overrides) -> Dataset:
     """BCI Competition IV 2a — 4-class motor imagery, 9 subjects."""
     defaults = dict(
         hf_id="braindecode/bcic2a",
+        revision="3585206f74463e787688a0f4afd12605c9be700c",
         n_classes=4,
         splitter=PredefinedSplitter(
             metadata_key="subject",
@@ -49,6 +51,7 @@ def bcic2020_3(**overrides) -> Dataset:
     """BCI Competition 2020-3 — 5-class imagined speech."""
     defaults = dict(
         hf_id="braindecode/bcic2020-3",
+        revision="8e2addefed44280be79b0c29bb09cc4c723cf18d",
         n_classes=5,
         montage_name="standard_1005",
         splitter=PredefinedSplitter(
@@ -66,6 +69,7 @@ def physionet(**overrides) -> Dataset:
     """PhysioNet Motor Imagery — 4 classes, 109 subjects."""
     defaults = dict(
         hf_id="braindecode/physionet",
+        revision="3c9cf8b242ca49550a43028bde3c6db3f1353b51",
         n_classes=4,
         splitter=PredefinedSplitter(
             metadata_key="subject",
@@ -82,6 +86,7 @@ def chbmit(**overrides) -> Dataset:
     """CHB-MIT seizure detection — 2 classes, 23 subjects."""
     defaults = dict(
         hf_id="braindecode/chbmit",
+        revision="3c7fc5537fb22686b2aafa8eb1a30d292c4464ab",
         n_classes=2,
         montage_name="standard_1005",
         splitter=PredefinedSplitter(
@@ -99,6 +104,7 @@ def faced(**overrides) -> Dataset:
     """FACED emotion recognition — 9 classes, 123 subjects."""
     defaults = dict(
         hf_id="braindecode/faced",
+        revision="18c73355fc71c055e92828384da85bb49bc6e3c0",
         n_classes=9,
         montage_name="standard_1005",
         splitter=PredefinedSplitter(
@@ -116,6 +122,7 @@ def isruc_sleep(**overrides) -> Dataset:
     """ISRUC sleep staging — 5 classes, ~60 usable subjects."""
     defaults = dict(
         hf_id="braindecode/isruc-sleep",
+        revision="beda68130bc3a84c8e6abf11c527d7c5ff941616",
         n_classes=5,
         montage_name="standard_1005",
         splitter=PredefinedSplitter(
@@ -143,6 +150,7 @@ def mdd_mumtaz2016(**overrides) -> Dataset:
     """MDD detection (Mumtaz 2016) — 2 classes."""
     defaults = dict(
         hf_id="braindecode/mdd_mumtaz2016",
+        revision="53cd9de1af5f09209d8a0f4fe20aa325c6a5fedb",
         n_classes=2,
         montage_name="standard_1005",
         splitter=PredefinedSplitter(
@@ -174,6 +182,7 @@ def seed_v(**overrides) -> Dataset:
     """SEED-V emotion recognition — 5 classes, session-based split."""
     defaults = dict(
         hf_id="braindecode/seed-v",
+        revision="b3498f4b045ade3093ecc646e2162057309479d3",
         n_classes=5,
         splitter=PredefinedSplitter(
             metadata_key="session",
@@ -190,6 +199,7 @@ def seed_vig(**overrides) -> Dataset:
     """SEED-VIG vigilance regression — continuous perclos target."""
     defaults = dict(
         hf_id="braindecode/seed-vig",
+        revision="896f1ad81375dc1aaf53f6e73d9d0c829be7bb40",
         n_classes=None,  # regression
         splitter=PredefinedSplitter(
             metadata_key="subject",
@@ -206,6 +216,7 @@ def tuab(**overrides) -> Dataset:
     """TUH Abnormal EEG — 2 classes (normal/abnormal), predefined train/eval split."""
     defaults = dict(
         hf_id="braindecode/tuab",
+        revision="7afcc5a00b50cac0f7e6e7f36928f7ccd8b91744",
         n_classes=2,
         splitter=PredefinedSplitter(
             metadata_key="train",
@@ -222,6 +233,7 @@ def tuev(**overrides) -> Dataset:
     """TUH EEG Events — 6 classes, predefined train/eval split."""
     defaults = dict(
         hf_id="braindecode/tuev",
+        revision="a16d2dcaf354a4c1fa1d31e1e591294802939868",
         n_classes=6,
         splitter=PredefinedSplitter(
             metadata_key="split",

open_eeg_bench/default_configs/datasets.py

@@ -21,7 +21,7 @@ classifiers = [
 ]
 
 dependencies = [
-    "braindecode[hub]>=1.4.0",
+    "braindecode[hub]>=1.6.1.dev184162698",
     "exca>=0.5.10",
     "peft>=0.14.0",
     "pydantic>=2.0",
@@ -43,9 +43,6 @@ Changelog = "https://github.com/braindecode/OpenEEGBench/blob/master/CHANGELOG.m
 benchopt = ["benchopt>=1.9"]
 dev = ["pytest", "pytest-cov"]
 
-[tool.uv.sources]
-braindecode = { git = "https://github.com/braindecode/braindecode.git" }
-
 [tool.setuptools.packages.find]
 where = ["."]
 include = ["open_eeg_bench*"]