--- library_name: transformers pipeline_tag: feature-extraction license: apache-2.0 language: - en tags: - eeg - neuroscience - foundation-model - masked-autoencoder - representation-learning - pytorch - custom_code --- # PRATIBHA-2, formerly known as "MANAS-2" PRATIBHA-2 (also called MANAS-2) is a transformer-based foundation model for EEG representation learning, developed by [MannasAI](https://huggingface.co/MannasAI). It combines **Raw-Band Hybrid (RBH)** masked autoencoding with **Constrained Reconstruction (ConRec)** to learn temporal and spectral features from EEG. ## Model Details PRATIBHA-2 has **92.8M parameters**, 22 transformer layers, and 512-dimensional embeddings. It processes EEG sampled at **200 Hz** using one-second patches with a 0.1-second overlap and spatiotemporal positional encoding. Channel count and signal length can vary. During pretraining, RBH uses waveform and spectral reconstruction, while ConRec regularizes local RMS-energy continuity in reconstructed waveforms. Pretraining uses unlabeled clinical EEG from internal sources, TUH EEG, and I-CARE. This release contains the encoder for feature extraction and fine-tuning. **Paper:** [MANAS-2: Constrained Reconstruction for EEG Foundation Models](https://arxiv.org/abs/2609.13717) ## Uses Install the dependencies: ```bash pip install "torch>=2.2" "transformers>=4.57,<5" "safetensors>=0.4" ``` Extract embeddings using the included electrode-position lookup: ```python import torch from transformers import AutoModel model = AutoModel.from_pretrained( "MannasAI/MANAS-2", trust_remote_code=True, ).eval() # Example: 10 seconds of EEG at 200 Hz. electrode_names = ["Fz", "Cz", "Pz"] eeg = torch.randn(1, 3, 2000) # [batch, channels, samples] eeg = (eeg - eeg.mean(-1, keepdim=True)) / ( eeg.std(-1, keepdim=True, correction=1) + 1e-6 ) positions = model.get_channel_positions(electrode_names) positions = positions.unsqueeze(0).expand(eeg.size(0), -1, -1) with torch.inference_mode(): features = model(eeg, positions) # [1, 11, 3, 512] embedding = features.mean(dim=(1, 2)) # [1, 512] ``` For real EEG, filter and resample to 200 Hz, then normalize each channel within each window as above. Electrode names must match the signal's channel order; the lookup returns coordinates in centimeters. The model expects preprocessed inputs and returns **[batch, patches, channels, 512]** features. ## Metrics PRATIBHA-2 has been evaluated using [NeuralBench](https://github.com/facebookresearch/neuroai/tree/main/neuralbench-repo). It outperforms all open source FMs currently included in the benchmark (as of 15/11/26), with an average rank across 34 tasks (NeuralBench's "core" tasks set) of 3.94. **Average Ranks** (lower is better) | Foundation model | Average rank ↓ | |---|---:| | **Pratibha-2** | **3.94** | | LaBraM | 4.94 | | REVE | 5.00 | | LUNA | 6.47 | | CBraMod | 6.76 | | BENDR | 10.71 | | BIOT | 12.32 | **Per-task performance** | Category | Task | Metric | Pratibha-2 | |---|---|---|---:| | Cognitive | Image | Top-5 accuracy (%) | 86.12 ± 1.64 | | Cognitive | Sentence | Top-5 accuracy (%) | 12.69 ± 0.60 | | Cognitive | Typing | Balanced accuracy (%) | 4.66 ± 0.27 | | Cognitive | Word | Top-5 accuracy (%) | 16.59 ± 0.83 | | BCI | c-VEP | Balanced accuracy (%) | 87.06 ± 0.73 | | BCI | Mental imagery | Balanced accuracy (%) | 27.54 ± 1.27 | | BCI | Motor execution | Balanced accuracy (%) | 55.12 ± 2.40 | | BCI | Motor imagery | Balanced accuracy (%) | 66.81 ± 0.20 | | BCI | P300 | Balanced accuracy (%) | 66.00 ± 0.38 | | BCI | SSVEP | Balanced accuracy (%) | 92.96 ± 0.71 | | Evoked Responses | Audiovisual | Balanced accuracy (%) | 43.69 ± 6.19 | | Evoked Responses | ERN | Balanced accuracy (%) | 78.14 ± 3.26 | | Evoked Responses | LRP | Balanced accuracy (%) | 78.69 ± 1.13 | | Evoked Responses | MMN | Balanced accuracy (%) | 58.05 ± 0.55 | | Evoked Responses | N170 | Balanced accuracy (%) | 71.35 ± 2.82 | | Evoked Responses | N2pc | Balanced accuracy (%) | 63.96 ± 0.67 | | Evoked Responses | N400 | Balanced accuracy (%) | 64.83 ± 1.52 | | Clinical | Clinical event | Macro F1 | 0.562 ± 0.024 | | Clinical | Dementia | Balanced accuracy (%) | 42.34 ± 6.03 | | Clinical | Depression | Balanced accuracy (%) | 88.37 ± 1.18 | | Clinical | Parkinson’s | Balanced accuracy (%) | 69.96 ± 0.49 | | Clinical | Pathology | Balanced accuracy (%) | 79.43 ± 0.79 | | Clinical | Schizophrenia | Balanced accuracy (%) | 67.27 ± 4.51 | | Clinical | Seizure | Balanced accuracy (%) | 70.89 ± 0.39 | | Internal State | **Emotion** | **Balanced accuracy (%)** | **30.21 ± 0.18** | | Internal State | Mental arithmetic | Balanced accuracy (%) | 73.54 ± 2.63 | | Internal State | Workload | Balanced accuracy (%) | 70.09 ± 1.80 | | Sleep | Sleep arousal | Balanced accuracy (%) | 58.02 ± 0.72 | | Sleep | Sleep stage | Balanced accuracy (%) | 74.23 ± 0.12 | | Phenotyping | Age | Pearson R | 0.706 ± 0.012 | | Phenotyping | Psychopathology | Pearson R | 0.089 ± 0.015 | | Phenotyping | Sex | Balanced accuracy (%) | 91.91 ± 0.38 | | Misc | Artifact | Macro F1 | 0.550 ± 0.006 | | Misc | Reaction time | Pearson R | 0.562 ± 0.004 | ## License Model weights and code are released under [Apache-2.0](LICENSE). ## Citation ```bibtex @misc{kulkarni2026PRATIBHA2, title={MANAS-2: Constrained Reconstruction for EEG Foundation Models}, author={Arvasu Kulkarni and Aditya Ray Mishra and Mahir Jain and Parshva Runwal and Lakshya Saini and Siddharth Panwar and Sandeep Singh}, year={2026}, eprint={2609.13717}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2609.13717} } ```