Upload README.md with huggingface_hub

README.md

@@ -5,68 +5,31 @@ tags:
 - brain-signals
 - neuroscience
 - reading
-- zuco
+- emmt
 - brain-to-text
 - cognitive-neuroscience
-license: mit
+- eye-tracking
+license: cc-by-4.0
 task_categories:
 - text-generation
-pretty_name: ZuCo EEG Reading Dataset
+pretty_name: EMMT EEG-to-Text Dataset
 size_categories:
-- 1K<n<10K
-dataset_info:
-  features:
-  - name: participant_id
-    dtype: string
-  - name: sentence_id
-    dtype: string
-  - name: sentence_text
-    dtype: string
-  - name: word_eeg_segments
-    list:
-      list:
-        list: float32
-  - name: num_words
-    dtype: int32
-  - name: eeg_shape
-    dtype: string
-  - name: has_gaze_alignment
-    dtype: bool
-  splits:
-  - name: train
-    num_bytes: 15090427
-    num_examples: 576
-  - name: validation
-    num_bytes: 3469583
-    num_examples: 128
-  - name: test
-    num_bytes: 4327666
-    num_examples: 160
-  download_size: 19869836
-  dataset_size: 22887676
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: validation
-    path: data/validation-*
-  - split: test
-    path: data/test-*
+- n<1K
 ---
 
-# ZuCo EEG Reading Dataset
+# EMMT EEG-to-Text Dataset
 
-This dataset contains preprocessed EEG brain signals recorded while participants read English sentences. It is derived from the [ZuCo (Zurich Cognitive Language Processing Corpus)](https://osf.io/q3zws/) dataset.
+This dataset contains preprocessed EEG brain signals recorded while participants read English sentences and translated them to Czech. It is derived from the **EMMT (Eyetracked Multi-Modal Translation)** corpus.
 
 ## Dataset Description
 
-- **Task**: EEG-to-Text (Brain-to-Text) Generation
-- **Modality**: EEG (Electroencephalography)
-- **Channels**: 4 (RAW_TP9, RAW_AF7, RAW_AF8, RAW_TP10)
+- **Task**: EEG-to-Text Generation (Brain-Computer Interface)
+- **Modality**: EEG (Electroencephalography) + Eye-tracking
+- **Channels**: 4 EEG channels (RAW_TP9, RAW_AF7, RAW_AF8, RAW_TP10)
 - **Sampling Rate**: 256 Hz
-- **Language**: English
-- **Participants**: Multiple subjects reading natural sentences
+- **Language**: English sentences (to be translated to Czech)
+- **Participants**: 21 participants
+- **Total Sentences**: 200 unique sentences
 
 ## Dataset Structure
 
@@ -74,175 +37,78 @@ This dataset contains preprocessed EEG brain signals recorded while participants
 
 | Split | Samples | Description |
 |-------|---------|-------------|
-| Train | 576 | 70% of participants for training |
-| Validation | 128 | 15% of participants for validation |
-| Test | 160 | 15% of participants for testing |
+| Train | 576 | Training set (70% of participants) |
+| Validation | 128 | Validation set (15% of participants) |
+| Test | 160 | Test set (15% of participants) |
 
-**Important**: Splits are done by **participant**, ensuring no data leakage. Different participants read different sentences, so validation and test sets contain both unseen participants and unseen sentence groups.
+**Important**: Splits are done by **participant** to prevent data leakage. Different participants may read the same sentences, ensuring the model learns to generalize across individuals.
 
 ### Data Fields
 
 Each sample contains:
 
-- **`participant`** (string): Participant ID (e.g., "P01", "P02")
-- **`sentence_id`** (string): Sentence identifier (e.g., "S001", "S002")
-- **`sentence`** (string): The actual text that was being read
+- **`participant_id`** (string): Anonymized participant identifier (e.g., "P01", "P02")
+- **`sentence_id`** (string): Sentence identifier (e.g., "S001", "S062")
+- **`sentence_text`** (string): The English sentence that was being read
 - **`word_eeg_segments`** (3D array): Word-level EEG data
   - Shape: `[num_words, time_samples, 4_channels]`
   - Each word has its own EEG segment extracted using gaze-timestamp alignment
-  - Time samples vary per word (typically 50-200 samples = 200-800ms)
+  - Time samples vary per word (typically 50-200 samples = 200-800ms @ 256Hz)
 - **`num_words`** (int): Number of words in the sentence
-- **`eeg_shape`** (string): Shape information (e.g., "15x128x4" = 15 words, 128 time samples, 4 channels)
+- **`eeg_shape`** (string): Human-readable shape info (e.g., "15words_x_128time_x_4channels")
+- **`has_gaze_alignment`** (bool): Whether gaze-based alignment was used
 
-### Preprocessing
+### Preprocessing Pipeline
 
-The EEG data has been preprocessed with:
+The EEG data has been preprocessed with standard neuroscience techniques:
 
-1. **Notch Filter**: 50Hz (removes power line noise)
-2. **Bandpass Filter**: 0.5-50Hz (keeps relevant brain frequencies)
-3. **Robust Scaling**: Normalization robust to eye blinks and artifacts
+1. **Notch Filter**: 50Hz (removes electrical power line noise)
+2. **Bandpass Filter**: 0.5-50Hz (keeps relevant brain frequencies, removes drift and high-freq noise)
+3. **Robust Scaling**: Normalization robust to eye blinks and movement artifacts
 4. **Gaze-Timestamp Alignment**: EEG segments aligned to actual word reading times using eye-tracking fixations
 
-## Usage
-
-### Loading the Dataset
-
-```python
-from datasets import load_dataset
-
-# Load the dataset
-dataset = load_dataset("sajjad5221/eeg2text-emmt-dataset")
-
-# Access splits
-train_data = dataset['train']
-val_data = dataset['validation']
-test_data = dataset['test']
-
-# Examine a sample
-sample = train_data[0]
-print(f"Participant: {sample['participant']}")
-print(f"Sentence: {sample['sentence']}")
-print(f"EEG shape: {sample['eeg_shape']}")
-print(f"Number of words: {sample['num_words']}")
-
-# Access word-level EEG data
-word_eeg = sample['word_eeg_segments']  # List of [time, channels] arrays
-first_word_eeg = word_eeg[0]  # EEG for first word: [time_samples, 4]
-print(f"First word EEG shape: {len(first_word_eeg)}x{len(first_word_eeg[0])}")
-```
-
-### Example: Training a Model
-
-```python
-import torch
-from torch.utils.data import DataLoader
-from datasets import load_dataset
-
-# Load dataset
-dataset = load_dataset("sajjad5221/eeg2text-emmt-dataset")
-
-# Convert to PyTorch tensors
-def collate_fn(batch):
-    # Your custom collation logic here
-    # Handle variable-length sequences
-    pass
-
-train_loader = DataLoader(
-    dataset['train'],
-    batch_size=8,
-    shuffle=True,
-    collate_fn=collate_fn
-)
-
-# Training loop
-for batch in train_loader:
-    eeg_data = batch['word_eeg_segments']
-    text = batch['sentence']
-    # Your training code here
-```
 
 ## Dataset Statistics
 
+### Overall Statistics
+- **Total Samples**: 864
+- **Unique Participants**: 21
+- **Unique Sentences**: 200
+- **Average Words per Sentence**: 11.0
+
 ### Training Set
-- Samples: 576
-- Unique participants: ~70% of total
-- Average words per sentence: ~10.9
+- **Samples**: 576
+- **Participants**: ~70% of total
+- **Purpose**: Model training
 
 ### Validation Set
-- Samples: 128
-- Unique participants: ~15% of total (different from training)
-- Used for: Model selection, hyperparameter tuning
+- **Samples**: 128
+- **Participants**: ~15% of total (different from training)
+- **Purpose**: Hyperparameter tuning, model selection
 
 ### Test Set
-- Samples: 160
-- Unique participants: ~15% of total (different from training & validation)
-- Used for: Final performance evaluation on unseen participants
+- **Samples**: 160
+- **Participants**: ~15% of total (different from training & validation)
+- **Purpose**: Final performance evaluation on unseen participants
 
 ## EEG Channel Information
 
-The 4 EEG channels capture different brain regions:
-
-| Channel | Location | Brain Region |
-|---------|----------|--------------|
-| TP9 | Left temporal-parietal | Language processing, memory |
-| AF7 | Left anterior-frontal | Attention, working memory |
-| AF8 | Right anterior-frontal | Attention, executive function |
-| TP10 | Right temporal-parietal | Visual processing, attention |
-
-## Citation
-
-If you use this dataset, please cite the original ZuCo paper:
-
-```bibtex
-@article{hollenstein2018zuco,
-  title={ZuCo, a simultaneous EEG and eye-tracking resource for natural sentence reading},
-  author={Hollenstein, Nora and Rotsztejn, Jonathan and Troendle, Marius and Pedroni, Andreas and Zhang, Ce and Langer, Nicolas},
-  journal={Scientific Data},
-  volume={5},
-  number={1},
-  pages={180291},
-  year={2018},
-  publisher={Nature Publishing Group}
-}
-```
-
-## Related Papers
-
-### ZuCo Dataset
-- [ZuCo 1.0](https://www.nature.com/articles/sdata2018291) - Original dataset paper
-- [ZuCo 2.0](https://www.nature.com/articles/s41597-022-01387-w) - Extended dataset
-
-### EEG-to-Text Models
-- [Brain2Text](https://arxiv.org/abs/2112.02685) - Neural decoding of text from brain signals
-- [DeWave](https://arxiv.org/abs/2309.14030) - Discrete EEG waves for text generation
-
-## Limitations
-
-- **Small vocabulary**: Limited to sentences in the ZuCo corpus
-- **Controlled setting**: Lab environment, not naturalistic reading
-- **Limited participants**: Dataset size constrained by number of participants
-- **Hardware dependency**: Recorded with specific EEG device (Muse headband)
-- **Individual differences**: Brain signals vary significantly across individuals
-
-## Ethics
-
-- All participants provided informed consent
-- Data has been anonymized (participant IDs are pseudonyms)
-- No sensitive personal information is included
-- Original study approved by ethics committee
-
-## License
+The 4 EEG channels capture different brain regions relevant to language processing:
 
-This dataset is released under the MIT License. Please refer to the original ZuCo dataset terms of use for any additional restrictions.
+| Channel | Location | Brain Region | Function |
+|---------|----------|--------------|----------|
+| **RAW_TP9** | Left temporal-parietal | Language processing | Semantic understanding, memory |
+| **RAW_AF7** | Left anterior-frontal | Executive function | Attention, working memory |
+| **RAW_AF8** | Right anterior-frontal | Executive function | Attention, cognitive control |
+| **RAW_TP10** | Right temporal-parietal | Visual-spatial | Visual processing, attention |
 
-## Contact
+## Original EMMT Experiment
 
-For questions or issues:
-- Open an issue on the [repository](https://huggingface.co/datasets/sajjad5221/eeg2text-emmt-dataset)
-- Refer to the [original ZuCo project](https://osf.io/q3zws/) for dataset-specific questions
+The data comes from the **EMMT (Eyetracked Multi-Modal Translation)** corpus, which recorded participants performing a multimodal translation task:
 
-## Acknowledgments
+1. **READ**: Read English sentence aloud
+2. **TRANSLATE**: Translate to Czech aloud
+3. **SEE**: View accompanying image
+4. **UPDATE**: Update or repeat translation
 
-- ZuCo dataset creators at University of Zurich
-- All participants who contributed their brain signals
-- Research groups advancing brain-computer interfaces
+This dataset contains EEG data from the **READ** stage, aligned to word-level using eye-tracking fixations.