---
license: mit
viewer: false
tags:
  - chemistry
  - raman-spectroscopy
  - deep-learning
  - self-supervised-learning
  - time-series
---

# GEMS: Multi-Source Raman Spectral Dataset

This repository hosts the comprehensive, multi-domain Raman spectral datasets utilized in the **GEMS** framework. 

The data is meticulously structured to support our proposed multi-stage training methodology, encompassing initial pre-training, contrastive learning, and diverse downstream fine-tuning tasks across different domains.

## Data Pre-processing & Format
All spectral data across the sub-directories have been uniformly pre-processed. They are **ready for direct model input** without requiring additional transformations:
* **Spectral Range:** 0 – 3500 cm⁻¹
* **Sequence Length:** Exactly 3,500 data points per spectrum.
* **Normalization:** Min-Max normalization has been applied to all samples.
* **Format:** Standard array formats ready for PyTorch/NumPy ingestion.

## Dataset Structure & Training Stages

The repository is organized into 7 distinct subsets. Each directory serves a highly specific role in the GEMS pipeline, ensuring continuous model optimization and rigorous experimental validation.

### Core Training Stages

| Directory | Stage | Primary Purpose |
| :--- | :--- | :--- |
| `QMe14S` | **Stage 1** | **Foundation Pre-training.** Used to train the initial foundational encoder. |
| `RRUFF_CL` | **Stage 2** | **Contrastive Learning.** Directly inherits the parameters from Stage 1 for continued training, aiming to build robust, generalized feature representations. |

### Downstream Fine-Tuning & Analytical Experiments
The remaining datasets leverage the pre-trained weights from Stage 2 for specific downstream applications and model evaluation:

| Directory | Experimental Focus / Task |
| :--- | :--- |
| `RRUFF_FT` | **Architecture & Hyperparameter Optimization.** Used as the benchmark to optimize the model's structural design and training configurations. |
| `Bacteria_ID` | **Few-Shot Learning Study.** Evaluates the model's generalization capabilities and performance stability when fine-tuned on highly limited annotated data. |
| `skincancer` | **Interpretability Analysis.** Investigates which specific spectral wavebands and features the deep learning model focuses on for medical diagnostics. |
| `Mutant_wheat` | **Downstream Fine-Tuning.** Agricultural domain classification task. |
| `microplastic` | **Downstream Fine-Tuning.** Environmental monitoring and material classification task. |

## Usage 

You can easily download the entire dataset or specific sub-directories using the official Hugging Face CLI:

```bash
# Download the entire dataset
hf download YourUsername/YourDatasetName --repo-type dataset --local-dir ./data

# Or download a specific subset (e.g., Stage 1 Pre-training data)
hf download YourUsername/YourDatasetName QMe14S/* --repo-type dataset --local-dir ./data/QMe14S