---
language:
- en
license: mit
task_categories:
- text-classification
task_ids:
- fact-checking
pretty_name: Project Sentinel – Semantic Filtering Dataset
---

# Project Sentinel – Semantic Filtering Dataset

## Dataset Summary

This dataset is designed to train and evaluate **lightweight semantic filtering models** for **Retrieval-Augmented Generation (RAG)** systems deployed in **resource-constrained / edge environments**.

It supports **binary classification** of `(query, candidate_sentence)` pairs into:

- **1 – Supporting Fact**: Factually useful evidence required to answer the query  
- **0 – Distractor**: Topically similar but factually insufficient or misleading text

The dataset was created to address the *hallucination-by-distraction* failure mode commonly observed in local RAG pipelines.

This dataset was used to fine-tune a **4-layer TinyBERT cross-encoder** for the *Project Sentinel* semantic gatekeeping module.

---

## Supported Tasks

- Semantic filtering
- Evidence verification
- Hallucination reduction in RAG
- Binary text classification

---

## Dataset Structure

The dataset consists of **sentence-level query–evidence pairs** derived from the **HotpotQA (Distractor) dataset**.

Each sample follows the structure:

| Field | Type | Description |
|-----|------|------------|
| `query` | string | Natural language question |
| `sentence` | string | Candidate sentence extracted from context |
| `label` | int | 1 = supporting fact, 0 = distractor |

---

## Dataset Splits

| Split | Samples |
|------|--------|
| Train | 69,101 |
| Validation | 7,006 |

The class distribution is intentionally **imbalanced**, reflecting real-world retrieval conditions where distractors vastly outnumber true evidence.

---

## Data Construction

### Source Dataset

- **HotpotQA – Distractor Setting**
- Multi-hop QA dataset with explicit annotations for supporting facts and distractor paragraphs

### Positive Samples

- Extracted exclusively from annotated *supporting fact sentences*
- Based on `(title, sentence index)` ground-truth supervision

### Negative Samples

- Non-annotated sentences from gold paragraphs *(hard negatives)*
- Sentences from distractor paragraphs intentionally selected for topical similarity

This strategy produces **adversarial negatives** closely matching real retrieval errors.

---

## Preprocessing Pipeline

- Removal of null or malformed samples
- Sentence-level deduplication
- Text normalization and whitespace cleanup
- Controlled negative sampling
- Label integrity verification

---

## Input Representation

The dataset is intended for **cross-encoder architectures**.

Model input format:

```
[CLS] query [SEP] sentence [SEP]
```

- Maximum sequence length: 512 tokens
- Designed for TinyBERT / MiniLM-style encoders

---

## Class Imbalance Handling

Due to natural imbalance, training is expected to use:

- **Weighted cross-entropy loss**
- Recall-preserving thresholds

This ensures evidence recall is prioritized over aggressive filtering.

---

## Intended Use

✔ Training semantic gatekeepers for RAG  
✔ Edge-deployed LLM pipelines  
✔ Hallucination suppression  
✔ Early-exit decision systems  

---

## Limitations

- Derived from Wikipedia-based QA (HotpotQA)
- English-only
- Sentence-level relevance only (not passage-level)

Performance may vary when applied to highly domain-specific corpora.

---

## Citation

If you use this dataset, please cite:

```
@article{salih2026sentinel,
  title={Project Sentinel: Lightweight Semantic Filtering for Edge RAG},
  author={Salih, El Mehdi and Ait El Mouden, Khaoula and Akchouch, Abdelhakim},
  year={2026}
}
```

---

## Contact

**El Mehdi Salih**  
Mohammed V University – Rabat  
Email: elmehdi_salih@um5.ac.ma