datause-train-v1

This is the training corpus for the model adapter (ai4data/datause-extraction-v1).

Unlike previous versions of the corpus that relied on choices-based JSON structures (extract_json), this dataset uses a pure entity-and-relation extraction schema (extract_entities + relation mapping). This avoids count prediction collapse in GLiNER2 while enabling multi-attribute extraction and joint span linking (including geographic scope).

Rationale and Context: Forced Displacement, Refugees, and FCV

Rationale for Creation

This dataset was created to train robust token-classification and relation-extraction models capable of auditing and mapping dataset usage across Fragility, Conflict, and Violence (FCV) settings and forced displacement contexts. Research on refugees, internally displaced persons (IDPs), and host communities is heavily reliant on diverse and scattered data sources—ranging from household surveys to program registries.

Tracking which datasets are being utilized, who is producing them, and how they are integrated into policy analysis helps international organizations (such as the World Bank and UNHCR), researchers, and funding bodies:

1. Monitor Data Investments: Quantify the academic and policy impact of dedicated data initiatives (e.g., those funded by the World Bank-UNHCR Joint Data Center on Forced Displacement).
2. Identify Data Gaps: Pinpoint regions or FCV research domains lacking primary microdata.
3. Avoid Research Duplication: Map existing analyses to avoid redundant data collection efforts in challenging, insecure areas.

Domain Coverage

The corpus spans multiple sectors and settings common to development economics and humanitarian research:

• Humanitarian Registries: UNHCR's proGRES database, registration rolls from national border/refugee agencies, and program databases.
• Displacement Tracking Systems: IOM's Displacement Tracking Matrix (DTM) reports and the Internal Displacement Monitoring Centre (IDMC) registries.
• Household Surveys in FCV Contexts: Living Standards Measurement Study (LSMS) surveys, Demographic and Health Surveys (DHS), Multiple Indicator Cluster Surveys (MICS), and specialized welfare monitoring surveys (e.g., SHINE, SESRE).
• Geospatial & Spatial Databases: Climate/weather indicators, conflict event databases (e.g., ACLED), and satellite camp imagery.

Splits

Split	File	Rows	Description
train	train.jsonl	13,291	Blended synthetic & augmented training instances
validation	validation.jsonl	651	Augmented validation/eval set

---

Dataset Statistics

Entity Distribution

Entity Type	Train Count	Eval Count	Description
name	33,345	1,908	Verbatim name or descriptive identifier of the data source
acronym	15,369	567	Verbatim abbreviation or acronym
producer	10,611	433	Publishing organization or author
timeframe	13,127	408	Year or time period covered
datatype	24,542	1,240	Verbatim data format (e.g. survey, census, database)
specificity	33,345	1,908	Categorization span: named / descriptive / vague
usage	33,345	1,908	Contextual role span: primary / supporting / background
geography	5,142	251	Verbatim country or geographic region linked to data

Relation Distribution

Relation Type	Train Count	Eval Count	Description
has_specificity	33,345	1,908	Links a name to its specificity classification
has_usage	33,345	1,908	Links a name to its usage classification
has_datatype	24,542	1,240	Links a name to its datatype attribute
has_acronym	15,369	567	Links a name to its acronym
has_timeframe	13,127	408	Links a name to its timeframe
has_producer	10,611	433	Links a name to its producer
has_geography	6,345	280	Links a name to its geography

---

Schema & Format

Every line is a JSON object prepended with a classification label prompt. For example:

{
  "input": "specificity: named | descriptive | vague usage: primary | supporting | background | We employ the 2006 round of the Additional Rural Incomes Survey and Rural Economic & Demographic Survey (ARIS-REDS) from the National Council of Applied Economic Research in India to examine household wealth dynamics.",
  "output": {
    "entities": {
      "name": [
        "2006 round of the Additional Rural Incomes Survey and Rural Economic & Demographic Survey (ARIS-REDS)"
      ],
      "acronym": ["ARIS-REDS"],
      "producer": ["National Council of Applied Economic Research"],
      "timeframe": ["2006"],
      "datatype": ["Survey"],
      "specificity": ["named"],
      "usage": ["primary"],
      "geography": ["India"]
    },
    "relations": [
      {
        "has_acronym": {
          "head": "2006 round of the Additional Rural Incomes Survey and Rural Economic & Demographic Survey (ARIS-REDS)",
          "tail": "ARIS-REDS"
        }
      },
      {
        "has_producer": {
          "head": "2006 round of the Additional Rural Incomes Survey and Rural Economic & Demographic Survey (ARIS-REDS)",
          "tail": "National Council of Applied Economic Research"
        }
      },
      {
        "has_timeframe": {
          "head": "2006 round of the Additional Rural Incomes Survey and Rural Economic & Demographic Survey (ARIS-REDS)",
          "tail": "2006"
        }
      },
      {
        "has_datatype": {
          "head": "2006 round of the Additional Rural Incomes Survey and Rural Economic & Demographic Survey (ARIS-REDS)",
          "tail": "Survey"
        }
      },
      {
        "has_specificity": {
          "head": "2006 round of the Additional Rural Incomes Survey and Rural Economic & Demographic Survey (ARIS-REDS)",
          "tail": "named"
        }
      },
      {
        "has_usage": {
          "head": "2006 round of the Additional Rural Incomes Survey and Rural Economic & Demographic Survey (ARIS-REDS)",
          "tail": "primary"
        }
      },
      {
        "has_geography": {
          "head": "2006 round of the Additional Rural Incomes Survey and Rural Economic & Demographic Survey (ARIS-REDS)",
          "tail": "India"
        }
      }
    ]
  }
}

Model Training

This dataset is compatible with training GLiNER2 LoRA adapters targeting the encoder, span representation, and relation classifier:

lora_target_modules = ["encoder", "span_rep", "classifier"]

Since the metadata and classifications are extracted as entities and mapped using relations, training does not require count_loss_weight or count predictors.

Loading the Dataset

from datasets import load_dataset

ds = load_dataset("ai4data/datause-train-v1")

train_data = ds["train"]
val_data   = ds["validation"]

Usage with ai4data

You can use the ai4data library to extract dataset mentions using the trained model adapter:

from ai4data.data_use import extract_from_text

text = (
    "To evaluate the impact of the refugee influx on local infrastructure, "
    "we analyze displacement trends using the 2023 UNHCR proGRES database "
    "and the IOM Displacement Tracking Matrix (DTM) in South Sudan."
)

result = extract_from_text(
    text,
    adapter_id="ai4data/datause-extraction-v1",
    include_confidence=True
)

# Access extracted dataset mentions
for dataset in result["datasets"]:
    print(f"Dataset Name: {dataset['mention_name']['text']}")
    print(f"Acronym: {dataset['acronym']['text']}")
    print(f"Producer: {dataset['producer']['text']}")
    print(f"Geography: {dataset['geography']['text']}")
    print(f"Specificity: {dataset['specificity_tag']['text']}")
    print(f"Usage Context: {dataset['usage_context']['text']}")