# 📗 SPECTER2 – Social Sciences Classifier (Binary Classification)

This model is a fine-tuned version of **allenai/specter2_base** for identifying whether a scientific publication belongs to the **Social Sciences** domain.

It achieves the following results on the evaluation set:

- Loss: 0.1382
- Accuracy: 0.9670
- F1 Micro: 0.9670
- F1 Macro: 0.9480
- F1 Weighted: 0.9670

## Model description

This model performs **binary document classification** and predicts whether a publication belongs to the **Social Sciences** domain.

The model was trained using title and abstract text from multiple openly available datasets with native disciplinary annotations, including:

- [MAG / SciDocs](https://github.com/allenai/scidocs)
- [Elsevier Open Access (ASJC subject areas)](https://researchcollaborations.elsevier.com/en/datasets/elsevier-oa-cc-by-corpus/)
- [ERC panel datasets (publications and funded projects)](https://huggingface.co/datasets/SIRIS-Lab/erc-classification-dataset)

Each dataset was converted into a common binary label indicating whether a document belongs to the Social Sciences according to mappings from the original classification systems. :contentReference[oaicite:0]{index=0}

**Key characteristics**

- Base model: `allenai/specter2_base`
- Task: binary document classification
- Labels:
  - `False` → Non-Social Sciences
  - `True` → Social Sciences
- Activation: softmax
- Loss: CrossEntropyLoss

## Intended uses

This model is suitable for:

- Identifying Social Sciences publications
- Research information systems
- Funding portfolio analysis
- Metadata enrichment
- Bibliometric analyses

The model accepts:

- title
- abstract
- title + abstract (recommended)

## Training data

Training data combines approximately **20,000** documents sampled from multiple sources:

- MAG/SciDocs
- ERC panel datasets (publications and projects)
- Elsevier Open Access publications

Each source provides its own disciplinary taxonomy. Categories corresponding to **Social Sciences** were mapped into a common binary classification problem. :contentReference[oaicite:1]{index=1}

## Training procedure

### Preprocessing

- Input text: `title + abstract`
- Maximum sequence length: **512 tokens**
- Tokenization using the SPECTER2 tokenizer

### Training hyperparameters

- learning_rate: 2e-5
- train_batch_size: 32
- eval_batch_size: 32
- num_epochs: 4
- max_length: 512
- optimizer: AdamW
- metric for best model: F1 Macro

## Evaluation results

| Metric | Value |
|--------|------:|
| Accuracy | 0.9670 |
| F1 Micro | 0.9670 |
| F1 Macro | 0.9480 |
| F1 Weighted | 0.9670 |

## Limitations

- The model predicts whether a publication belongs to the **Social Sciences** domain only.
- It does not distinguish between individual Social Sciences disciplines.
- Labels are derived from mappings between different disciplinary taxonomies and should be interpreted as high-level domain assignments.

## Framework versions

- Transformers 4.57.1
- PyTorch 2.8.0
- Datasets 3.6.0
- Tokenizers 0.22.1