File size: 5,104 Bytes
ef1de23 a3128bb ef1de23 a3128bb ef1de23 a3128bb ef1de23 a3128bb ef1de23 a3128bb a003b4d a3128bb a003b4d a3128bb a003b4d a3128bb a003b4d a3128bb ef1de23 a3128bb ef1de23 a3128bb a003b4d | 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 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 | ---
license: cc-by-4.0
language:
- en
tags:
- ner
- span-extraction
- token-classification
- data-mentions
- information-extraction
- displacement
task_categories:
- token-classification
configs:
- config_name: gliner_displacement
data_files:
- split: train
path: gliner_displacement_train.jsonl
- split: val
path: gliner_displacement_val.jsonl
- split: holdout
path: gliner_displacement_holdout.jsonl
- config_name: bio_displacement
data_files:
- split: train
path: bio_displacement_train.jsonl
- split: val
path: bio_displacement_val.jsonl
- split: holdout
path: bio_displacement_holdout.jsonl
- config_name: readable_displacement
data_files:
- split: train
path: readable_displacement_train.jsonl
- split: val
path: readable_displacement_val.jsonl
- split: holdout
path: readable_displacement_holdout.jsonl
---
# Data-Use Mentions — Displacement View
Derived from [`rafmacalaba/data-use-mentions`](https://huggingface.co/datasets/rafmacalaba/data-use-mentions) (originals untouched). Every source span keeps its label verdict, so the kept-vs-dropped split is explicitly trackable and auditable in every row — nothing is silently removed.
- **kept → `DATA_MENTION`**: evidential and declaration data-use mentions (the positive label).
- **dropped → `NON_MENTION`**: non-mentions and junk (the negative label).
## Why “displacement”
Non-mention/junk spans are *displaced* from the positive label set — so training sees them as negatives — but stay in the data with their verdict attached, rather than being deleted from the record.
## Configurations
| config | fields | purpose |
|---|---|---|
| `gliner_displacement` | `tokenized_text`, `corpus`, `origin`, `language`, `ner`, `spans` | GLiNER-style span extraction; `ner` lists only kept `DATA_MENTION` spans, `spans` carries the per-span `uid`, `tier`, and `source` for full auditability |
| `bio_displacement` | `tokens`, `ner_tags`, `spans`, `corpus`, `origin`, `language` | token-level BIO with two explicit classes (`DATA_MENTION`, `NON_MENTION`), overlap-safe longest-first; `spans` is the readable reconstruction of the tags |
| `readable_displacement` | `text`, `spans`, `corpus`, `origin`, `origin_model`, `language` | human-readable: full window `text` with character-offset `spans` (`{start, end, text, label, tier, source}`) |
## Label scheme
| label | meaning |
|---|---|
| `DATA_MENTION` | kept span — evidential or declaration data-use mention (positive) |
| `NON_MENTION` | dropped span — non-mention or junk (negative) |
## Verdict provenance
- Luna judgments win for any span they cover (val/holdout, plus the judged train subset).
- Unjudged train spans use the v3 classifier's predicted class (`train_v3_remaining_scores.jsonl`).
- A span with no resolvable id stays tagged `DATA_MENTION` (asymmetric keep: never risk dropping a positive).
- Same-surface conflicts keep when any positive verdict is present.
- A Luna review pass (`refine_band.py`, label-blind) adjudicated the ~6.3k highest-risk kept spans across all splits; spans it confirmed T3/junk flipped to `NON_MENTION` (source `luna_review`), confirmed T1/T2 stayed.
## Splits and sizes
Total: **122,864 rows**, **148,584 spans** (117,380 kept, 31,204 dropped, 1,149 same-surface conflicts).
| split | rows | spans | kept (`DATA_MENTION`) | dropped (`NON_MENTION`) | conflicts |
|---|---:|---:|---:|---:|---:|
| train | 85,003 | 103,230 | 81,995 | 21,235 | 758 |
| val | 19,152 | 23,302 | 18,141 | 5,161 | 229 |
| holdout | 18,709 | 22,052 | 17,244 | 4,808 | 162 |
## Source corpora
Rows come from the `fcv` and `prwp` corpora with five source origins: `fcv_pads_east_africa`, `general_prwp`, `reliefweb`, `refugee_pads`, `jdc_operational`. This dataset contains only real (non-synthetic) rows; `origin_model` is `""` for every row.
## Usage
```python
from datasets import load_dataset
ds = load_dataset("rafmacalaba/datause-displacement",
"readable_displacement", split="train")
```
The `origin` field is normalized (legacy suffixes stripped).
`{"holdout": {"by_source": {"luna": 21409}, "conflict_spans": 162, "dropped_T3junk": 4808, "kept_T1T2": 17244, "review_drop_flips": 159, "review_keep_flips": 0, "rows": 18709, "spans": 22052, "tiers": {"junk": 299, "tier1_evidential": 13326, "tier2_declaration": 3918, "tier3_nonmention": 4509}}, "train": {"by_source": {"classifier": 70385, "luna": 26779}, "conflict_spans": 758, "dropped_T3junk": 21235, "kept_T1T2": 81995, "review_drop_flips": 1983, "review_keep_flips": 0, "rows": 85003, "spans": 103230, "tiers": {"junk": 1758, "tier1_evidential": 63880, "tier2_declaration": 18115, "tier3_nonmention": 19477}}, "val": {"by_source": {"luna": 22552}, "conflict_spans": 229, "dropped_T3junk": 5161, "kept_T1T2": 18141, "review_drop_flips": 172, "review_keep_flips": 0, "rows": 19152, "spans": 23302, "tiers": {"junk": 299, "tier1_evidential": 14169, "tier2_declaration": 3972, "tier3_nonmention": 4862}}}`
|