File size: 2,326 Bytes
7dcd9f7 2def799 7dcd9f7 2def799 7dcd9f7 2def799 7dcd9f7 2def799 f217a79 7dcd9f7 2def799 4494be2 2def799 4494be2 2def799 7dcd9f7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 | ---
license: cc-by-4.0
language: [en]
tags:
- data-use
- provenance
- usage-impact
- world-bank
- fcv
- named-entity-recognition
configs:
- config_name: default
data_files:
- split: train
path: train.jsonl
---
# fcv-extractions-meta
Data-use mention extractions from the World Bank Fragility, Conflict and Violence (FCV) document corpus, enriched by the fine-tuned `rafmacalaba/lfm2.5-350M-datause-multitask` model (a LoRA SFT of `LiquidAI/LFM2.5-350M`).
## Shape
long — one row per mention; `provenance` and `usage_impact` are flattened to top-level columns (`data_type`, `usage_action`, `impact_label`, `usage_summary`, `producer`, `year`, `geography`, `acronym`).
## Configs
| config | rows |
| --- | --- |
| default | 33,471 |
## Fields
- **Document trace**: `corpus_id`, `config`, `pdf_url`, `title`, `page`, `chunk`.
- **Mention span**: `mention` (long) or `entities[].text` (nested), `label` (`NAMED_DATA` / `DESCRIPTIVE_DATA` / `VAGUE_DATA`), `score`, `start`, `end`.
- **Document metadata** (when built with `--with-metadata`): `project_id`, `doc_id`, `document_type`, `country`, `sector`, `language`, `disclosure_status`, `document_date`, `region`, `topics`, `themes`, `authors`, `abstract`, and more, joined from `rafmacalaba/fcv-corpus`.
### provenance
`producer`, `year`, `geography`, `acronym` — verbatim substrings of the source text; keys are omitted when the model finds nothing.
### usage_impact
- `data_type`: survey, administrative, indicator, database, geospatial, report, other
- `usage_action`: source, analyze, curate, validate, inform
- `impact_label`: none, evidence-base, policy, uptake
- `usage_summary`: one-sentence free text describing how the author uses the data.
## Attribution coverage
33,328 of 33,471 extracted entities matched an inference row (99.6%). Entities without a match carry only the raw extraction fields (no `provenance` / `usage_impact`).
## How it was built
1. `data_use/run.py` — GLiNER extraction of data mentions over chunked document pages.
2. `data_use/infer_multitask.py` — LFM2.5-350M multitask model predicts provenance + usage/impact per mention.
3. `data_use/attribute.py` — joins the inference back onto the extraction chunks on `(corpus_id, page, chunk, start, end)` and attaches corpus document labels.
|