README.md

---
base_model: fastino/gliner2-large-v1
library_name: peft
tags:
- base_model:adapter:fastino/gliner2-large-v1
- lora
- transformers
- gliner2
- dataset-extraction
- data-use
---

# GLiNER2 Data Use Extraction Adapter (v2)

This is a fine-tuned LoRA adapter for `fastino/gliner2-large-v1` trained to extract datasets, data mentions, and their relations from academic papers, research, and reports (with a focus on World Bank/UNHCR documents).

- **Repository:** [https://github.com/rafmacalaba/monitoring_of_datause](https://github.com/rafmacalaba/monitoring_of_datause)
- **Base Model:** `fastino/gliner2-large-v1`
- **Adapter ID:** `ai4data/datause-extraction-v2`

---

## How to Get Started with the Model

It is **highly recommended** to use this model through the official **`ai4data`** Python library wrapper. The library automatically handles:
- **Markdown-aware chunking** (respecting model context limits).
- **Character offset index adjustment** across multiple chunk pages.
- **Greedy overlap resolution** and text normalization.
- **Pre-filtering pre-classifiers** to skip non-data pages.
- **Deduplication** and acronym matching.

### 1. Installation

Clone and install the repository:

```bash
git clone <repository-url>
cd monitoring_of_datause
uv sync
```

### 2. Python Usage

To extract dataset mentions and their attributes (like timeframe, producer, and acronyms):

```python
from ai4data import extract_from_text, extract_from_document

text = """Our analysis uses the 2022 Demographic and Health Survey (DHS) conducted by
the National Statistics Office. We complement this with administrative systems, but
only the DHS is used in the empirical models."""

# Extract from raw text
results = extract_from_text(text)
print(results["datasets"])

# Extract from a PDF document
pdf_results = extract_from_document("report.pdf", pages=[0, 1, 2])
print(pdf_results)
```

---

## Model Schema & Response Structure

The model extracts up to 7 attributes per data mention. When querying via `ai4data`, each extracted entity in the `"datasets"` list has the following structure:

```json
{
  "mention_name": {
    "text": "Demographic and Health Survey",
    "confidence": 0.9999,
    "start": 23,
    "end": 52
  },
  "specificity_tag": {
    "text": "named",
    "confidence": 0.9999,
    "start": 23,
    "end": 52
  },
  "usage_context": {
    "text": "primary",
    "confidence": 0.9999,
    "start": 23,
    "end": 52
  },
  "typology_tag": {
    "text": "survey",
    "confidence": 0.9999,
    "start": 23,
    "end": 52
  },
  "acronym": {
    "text": "DHS",
    "confidence": 0.9996,
    "start": 54,
    "end": 57
  },
  "producer": {
    "text": "National Statistics Office",
    "confidence": 0.9999,
    "start": 72,
    "end": 98
  },
  "reference_year": {
    "text": "2022",
    "confidence": 0.9999,
    "start": 18,
    "end": 22
  },
  "is_used": {
    "text": "True",
    "confidence": 0.9999,
    "start": 23,
    "end": 52
  },
  "geography": {
    "text": "",
    "confidence": 0.9999,
    "start": 23,
    "end": 52
  }
}
```

---

## Annotation Guidelines & What Counts as a Data Mention

### Specificity Taxonomy
- **`named`**: A specific, citable dataset (e.g., `"DHS 2020"`, `"World Development Indicators"`, `"Ghana Living Standards Survey (GLSS)"`)
- **`descriptive`**: A general category of data, not a specific named dataset (e.g., `"household survey data"`, `"administrative records"`, `"panel data on firms"`)
- **`vague`**: An indirect or ambiguous reference (e.g., `"available data"`, `"our dataset"`, `"the data used in this study"`)

### Usage Context
- **`primary`**: Core data driving the main analysis in the report.
- **`supporting`**: Secondary data used to validate, calibrate, or provide robustness checks.
- **`background`**: Mentioned in passing, in a literature review, or as historical context.