FCV data-use paper corpus (6 configs)

README.md

@@ -48,10 +48,6 @@ configs:
   data_files:
   - split: train
     path: "synthetic/train.jsonl"
-- config_name: documents
-  data_files:
-  - split: all
-    path: "documents/all.jsonl"
 - config_name: sources
   data_files:
   - split: all
@@ -67,7 +63,7 @@ Extraction Framework"** (Macalaba, Solatorio, Brock).
 Three-model swarm over the forced-displacement / FCV corpus: dataset-mention
 extraction (Call 1), relation extraction (Call 1b, 5 types), and impact/usage
 classification (Call 2, 3 tasks) -- plus the seed-based synthetic training data,
-the per-document active-learning corpus, and the per-source original extractions.
+the PRWP training subset, and the per-source original extractions.
 
 All splits are published verbatim exactly as used in the paper.
 
@@ -75,11 +71,10 @@ All splits are published verbatim exactly as used in the paper.
 
 | Config | Splits (records) |
 |---|---|
-| `extraction` | train: 2473, validation: 415, test: 1706 |
+| `extraction` | train: 5183, validation: 415, test: 1706 |
 | `relation` | train: 640, validation: 91, holdout: 21 |
 | `impact` | train: 604, validation: 151, holdout: 84 |
 | `synthetic` | train: 1794 |
-| `documents` | all: 4435 |
 | `sources` | all: 5445 |
 
 ## Schemas
@@ -87,20 +82,19 @@ All splits are published verbatim exactly as used in the paper.
 - `extraction` / `synthetic` / `sources`: `{"input", "output": {"entities":
   {"named_data", "descriptive_data", "vague_data"}, "entity_descriptions"}}`;
   `extraction/test` additionally carries the model's `predictions`.
+- `extraction` rows carry a `meta` field tagging provenance: `FCV`
+  (forced-displacement / FCV corpus) or `PRWP` (Policy Research Working Papers
+  -- manually annotated + reviewed/validated). `train` = FCV (2,473) + PRWP
+  (2,710); `validation` and `test` are FCV-only.
 - `relation`: `{"input", "output": {"entities": {"named_data",
   "organization"}, "relations": [{"<type>": {"head", "tail"}}]}, "_meta"}`
   with types `has_organization`, `used_by`, `has_acronym`, `has_timeframe`,
   `has_geography`.
 - `impact`: `{"input", "output": {"classifications": [{"task", "labels",
   "true_label"}]}}` with tasks `purpose_action`, `usage`, `typology`.
-- `documents`: one row per validated chunk with `document_id`,
-  `corpus_category`, `split`, `active_learning`, `page_index`, `page_number`,
-  `text`, `mentions` (verdict, relations, impact fields). The 57 active-learning
-  documents (32 UNHCR + 14 PRWP + 11 PAD) are flagged `active_learning: true`
-  with their paper Table-1 `split` (29 train / 28 holdout).
 - `sources`: adds a `source` column (`prwp`, `reliefweb`, `seis`, `esmap`,
-  `refugee_pads`) plus `document_id` and `active_learning`; the 57 AL docs are
-  emitted as flat-NER rows (invalid mentions filtered by verdict) flagged
+  `refugee_pads`) plus `document_id` and `active_learning`; the 57 active-learning
+  docs are emitted as flat-NER rows (invalid mentions filtered by verdict) flagged
   `active_learning: true`.
 
 ## Citation