|
|
--- |
|
|
license: apache-2.0 |
|
|
tags: |
|
|
- factuality |
|
|
- grounding |
|
|
- llm-evaluation |
|
|
- text |
|
|
- preprocessing |
|
|
--- |
|
|
--- |
|
|
|
|
|
|
|
|
# Facts Grounding Processed |
|
|
|
|
|
This dataset is a **processed and reformatted version** of [`google/FACTS-grounding-public`](https://huggingface.co/datasets/google/FACTS-grounding-public), designed for **LLM factuality and grounding evaluation**. |
|
|
It has been cleaned, enriched with additional metadata, and split into **train** and **validation** sets. |
|
|
|
|
|
--- |
|
|
|
|
|
## Dataset Summary |
|
|
|
|
|
The dataset contains prompts, context documents, and target answers that challenge models to stay grounded in provided context rather than hallucinating. |
|
|
Processing steps added extra features like: |
|
|
- `prompt` – consolidated instruction + user request + context |
|
|
- `has_url_in_context` – boolean flag for URLs in context |
|
|
- `len_system`, `len_user`, `len_context` – token/word length statistics |
|
|
- `row_id` – unique identifier for tracking |
|
|
|
|
|
--- |
|
|
|
|
|
## Dataset Structure |
|
|
|
|
|
**Splits:** |
|
|
- **train** – 688 rows |
|
|
- **validation** – 172 rows |
|
|
|
|
|
**Features:** |
|
|
| Feature | Type | Description | |
|
|
|-----------------------|---------|-------------| |
|
|
| `system_instruction` | string | System-level instruction template | |
|
|
| `user_request` | string | User query or prompt | |
|
|
| `context_document` | string | Provided factual context | |
|
|
| `full_prompt` | string | Original concatenated instruction + context | |
|
|
| `prompt` | string | Processed consolidated prompt | |
|
|
| `has_url_in_context` | bool | Whether the context contains a URL | |
|
|
| `len_system` | int64 | Length of system instruction | |
|
|
| `len_user` | int64 | Length of user request | |
|
|
| `len_context` | int64 | Length of context document | |
|
|
| `target` | string | Grounded factual answer | |
|
|
| `row_id` | int64 | Unique row identifier | |
|
|
|
|
|
--- |
|
|
|
|
|
## Processing |
|
|
|
|
|
The preprocessing was done in Python with the Hugging Face `datasets` library: |
|
|
1. Loaded the original dataset (`google/FACTS-grounding-public`) |
|
|
2. Added custom features (URL detection, length counts) |
|
|
3. Generated combined prompt field |
|
|
4. Split into train (80%) and validation (20%) |
|
|
5. Exported as Arrow and JSONL formats |
|
|
|
|
|
The preprocessing script is included in the repository for reproducibility. |
|
|
|
|
|
--- |
|
|
|
|
|
## Intended Uses |
|
|
|
|
|
- **LLM grounding & hallucination evaluation** |
|
|
- Prompt engineering experiments |
|
|
- Fine-tuning or zero/few-shot evaluation pipelines |
|
|
- Data formatting utilities |
|
|
|
|
|
--- |
|
|
|
|
|
## How to Use |
|
|
|
|
|
```python |
|
|
from datasets import load_dataset |
|
|
|
|
|
dataset = load_dataset("GenAIDevTOProd/facts-grounding-processed") |
|
|
|
|
|
print(dataset["train"][0]) |
|
|
|
|
|
License: |
|
|
This processed dataset is licensed under Apache 2.0. |
|
|
The original data source is google/FACTS-grounding-public. |
