README.md

---
dataset_info:
  features:
  - name: data_source
    dtype: string
  - name: tokens
    list: string
  - name: tags
    list: string
  splits:
  - name: train
    num_bytes: 4506330
    num_examples: 8409
  - name: validation
    num_bytes: 739231
    num_examples: 1407
  - name: test
    num_bytes: 624990
    num_examples: 1295
  download_size: 1032656
  dataset_size: 5870551
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: cc-by-4.0
task_categories:
- token-classification
language:
- en
pretty_name: CeLLaTe-NER-2Class

tags:
- biomedical
- biomedical-ner
- named-entity-recognition
- cell-line
- cell-type
- tissue
- life-sciences
- transformers
- token-classification
---

# CeLLaTe-NER 2-Class

## Dataset Summary

**CeLLaTe-NER 2-Class** is a biomedical Named Entity Recognition (NER) dataset developed for recognising two biologically relevant entity types:

- **CellLine**
- **Cell_Tissue** (a merged entity combining the original *CellType* and *Tissue* labels)

The dataset was created as part of an iterative model development and error analysis cycle for biomedical NER. It is derived from an earlier three-class annotation scheme consisting of:

- CellLine
- CellType
- Tissue

During experimentation, qualitative and quantitative error analysis consistently showed that **CellType** and **Tissue** frequently occurred in highly overlapping linguistic contexts, making them difficult for models to distinguish reliably. Furthermore, **Tissue** was consistently the lowest-performing entity class across multiple experiments.

To address these observations, the label schema was revised by merging the **CellType** and **Tissue** entities into a single **Cell_Tissue** class. This simplified label space is intended to reduce label ambiguity and provides a more robust benchmark for developing biomedical NER models focused on cell-related entity recognition.

The dataset is intended primarily for training and evaluating transformer-based biomedical NER models.

---

# Motivation

Biomedical entity recognition often suffers from semantic ambiguity between closely related biological concepts.

For example:

- cell types frequently appear within tissue descriptions;
- tissues may be used adjectivally to describe cells;
- many biomedical publications refer to both concepts interchangeably depending on experimental context.

During development of the original CeLLaTe models, extensive model error analysis revealed that:

- the majority of misclassifications occurred between **CellType** and **Tissue**;
- Tissue consistently produced the lowest F1-score;
- many annotation boundaries and contextual cues overlapped between these entity types.

Rather than attempting to force increasingly complex models to distinguish these highly correlated concepts, this version of the dataset adopts a simplified annotation strategy that combines both entities into a unified **Cell_Tissue** category.

This allows models to focus on distinguishing biologically meaningful categories while reducing label ambiguity.

---

# Supported Tasks

This dataset supports:

- Biomedical Named Entity Recognition (NER)
- Token Classification
- Domain Adaptation
- Continual Learning
- Biomedical Language Model Fine-tuning
- Error Analysis of Biomedical NER Systems

---

# Label Schema

The dataset contains two entity classes.

| Label | Description |
|--------|-------------|
| **CellLine** | Established cell lines used in biomedical research. |
| **Cell_Tissue** | Combined entity representing both Cell Types and Tissue mentions. |

Example:

```
HeLa          B-CellLine
cells         I-CellLine
...

human         B-Cell_Tissue
fibroblasts   I-Cell_Tissue
...
```

---

# Why merge CellType and Tissue?

The original annotation schema consisted of three entity types:

- CellLine
- CellType
- Tissue

However, model evaluation revealed several recurring issues:

- significant contextual overlap between CellType and Tissue;
- high confusion rates between these labels;
- Tissue consistently exhibited the lowest predictive performance;
- many prediction errors were attributable to semantic similarity rather than model capacity.

The revised schema therefore combines CellType and Tissue into a unified **Cell_Tissue** label, producing a cleaner learning objective while preserving the biological information required for many downstream applications.

---

# Dataset Creation

This dataset was created through an iterative annotation and model development workflow.

The development process involved:

1. Training biomedical NER models using the original three-class schema.
2. Performing quantitative evaluation using standard NER metrics.
3. Conducting detailed qualitative error analysis.
4. Identifying systematic confusion between CellType and Tissue.
5. Revising the label schema.
6. Re-generating using the new two-class label space.

The resulting dataset represents an improved version of the original corpus informed by empirical model behaviour.

---

# Data Format

The dataset follows the standard token classification format.

Typical fields include:

| Column | Description |
|---------|-------------|
| tokens | Tokenized sentence |
| ner_tags | BIO entity labels |

Example:

```python
{
    "tokens": [
        "HeLa",
        "cells",
        "were",
        "cultured"
    ],
    "ner_tags": [
        "B-CellLine",
        "I-CellLine",
        "O",
        "O"
    ]
}
```

---

# Intended Uses

The dataset is intended for:

- Biomedical NER research
- Transformer model fine-tuning
- Domain-specific language model evaluation
- Continual learning experiments
- Error analysis
- Benchmarking biomedical token classification systems

---

# Models

Models trained using this dataset are available in the accompanying model repository.

Model names follow the naming convention:

```
CeLLaTe-ner-2class-...
```

These models demonstrate the effectiveness of the simplified annotation schema for biomedical NER.

---

# Evaluation

Typical evaluation metrics include:

- Precision
- Recall
- F1-score
- Entity-level Precision/Recall/F1
- Confusion Matrix
- Per-class Performance

Entity-level evaluation using the `seqeval` metric is recommended.

---

# Limitations

Although merging CellType and Tissue improves modelling performance, it also reduces the granularity of the annotations.

Researchers interested in distinguishing between these two concepts should instead use the original three-class dataset.

As with most biomedical corpora, the dataset may also inherit:

- publication bias;
- annotation ambiguity;
- terminology variation across biomedical subdomains.

---

# Citation

If you use this dataset, please cite:

```bibtex
@dataset{cellate_ner_2class,
  title={CeLLaTe-NER 2-Class},
  author={...},
  year={2026},
  publisher={Hugging Face}
}
```

---

# Acknowledgements

This dataset was developed as part of an iterative biomedical Named Entity Recognition research project focused on improving recognition of cell-related entities through data-centric model development and systematic error analysis.

The revised annotation schema was motivated by empirical observations from model evaluation, demonstrating the importance of combining qualitative error analysis with dataset refinement to improve downstream model performance.