rafmacalaba commited on
Commit
9369c2e
·
verified ·
1 Parent(s): 8c13690

raw-extraction build (model predictions, no judge)

Browse files
.gitattributes CHANGED
@@ -58,3 +58,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
 
 
 
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
61
+ gliner2_train.jsonl filter=lfs diff=lfs merge=lfs -text
62
+ gliner_train.jsonl filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ task_categories:
3
+ - token-classification
4
+ tags:
5
+ - ner
6
+ - span-extraction
7
+ - data-mentions
8
+ - economics
9
+ license: cc-by-4.0
10
+ configs:
11
+ - config_name: gliner
12
+ data_files:
13
+ - split: train
14
+ path: gliner_train.jsonl
15
+ - split: val
16
+ path: gliner_val.jsonl
17
+ - split: holdout
18
+ path: gliner_holdout.jsonl
19
+ - config_name: gliner2
20
+ data_files:
21
+ - split: train
22
+ path: gliner2_train.jsonl
23
+ - split: val
24
+ path: gliner2_val.jsonl
25
+ - split: holdout
26
+ path: gliner2_holdout.jsonl
27
+ ---
28
+
29
+ # datause-agents-v3 — agent-annotated data-use mentions
30
+
31
+ **AGENT-LABELED. None of these spans is an owner ruling.** Every entity was admitted by the
32
+ annotating agent under the doctrine rules (`C1`/`C2`/`S4`), with `owner_reviewed: false` on
33
+ every row. A model following the owner's guide agrees with the owner on ~71% of a fresh
34
+ sample, so treat holdout numbers from this split as *agreement with the agent*, not accuracy.
35
+
36
+ ## Labels
37
+
38
+ - `NAMED_DATA` — a proper name, title, or acronym of a specific data source
39
+ - `DESCRIPTIVE_DATA` — a source described in words but not named
40
+ - `VAGUE_DATA` — generic data wording with no identifiable source
41
+
42
+ ## Provenance
43
+
44
+ - bundle: `agent_annotations_v3` (`annotated_by: agent`, `owner_reviewed: false`)
45
+ - lanes: origin_strat_1, origin_strat_2, origin_strat_3, umar_pads, umar_pads_500
46
+ - carve: `agents-v3` (origin-stratified) — document-disjoint, so no
47
+ two passages of one document straddle train and eval, and stratified by origin, so each
48
+ split holds each corpus family's exact share
49
+ - guard: 77 eval-text and 253 duplicate-text passages excluded;
50
+ 128 eval documents checked, 0 leaks
51
+ - **not evaluable**: jdc_operational have no documents in val/holdout (too few to stratify), so they are train-only
52
+ and no split metric says anything about them
53
+ - labels: `labels.json`, per-split counts and per-origin document counts: `split_stats.json`
54
+
55
+ ## Configs
56
+
57
+ - `gliner` — `{"tokenized_text": [...], "ner": [[start, end, LABEL], ...]}` (`ner` is inclusive on both ends)
58
+ - `gliner2` — `{"input": "...", "output": {"entities": {grade_key: [span strings]}}}`
59
+
60
+ ## Splits
61
+
62
+ 21,073 passages · 16,167 spans · 12,831 documents, document-disjoint.
63
+
64
+ | split | rows | spans | documents | NAMED | DESCRIPTIVE | VAGUE |
65
+ | --- | ---: | ---: | ---: | ---: | ---: | ---: |
66
+ | train | 16,843 | 12,891 | 10,265 | 9,723 | 3,142 | 26 |
67
+ | val | 2,122 | 1,665 | 1,283 | 1,212 | 451 | 2 |
68
+ | holdout | 2,108 | 1,611 | 1,283 | 1,225 | 383 | 3 |
69
+
70
+ ## Usage
71
+
72
+ ```bash
73
+ uv run python training/finetune_gliner.py \
74
+ --dataset rafmacalaba/datause-agents-v3 --config gliner \
75
+ --epochs 5 --batch-size 16 --output-dir models/gliner_agents_v3
76
+ ```
77
+
78
+ ```python
79
+ from datasets import load_dataset
80
+ ds = load_dataset("rafmacalaba/datause-agents-v3", "gliner") # ds['train'] / ds['val'] / ds['holdout']
81
+ ```
gliner2_holdout.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
gliner2_train.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bab49a8c335a1ef44ae3aaf2b51180943a9f4afd6979683ae720cb357c74908a
3
+ size 33689687
gliner2_val.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
gliner_holdout.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
gliner_train.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8f886289d7f40ca7c269806eb017f2e9226e09e245fbc40fffe7f79ff8304d46
3
+ size 48418084
gliner_val.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
labels.json ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "gliner": [
3
+ "NAMED_DATA",
4
+ "DESCRIPTIVE_DATA",
5
+ "VAGUE_DATA"
6
+ ],
7
+ "gliner2": [
8
+ "named_data",
9
+ "descriptive_data",
10
+ "vague_data"
11
+ ]
12
+ }
split_stats.json ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "provenance": {
3
+ "datasets": [
4
+ "origin_strat_1",
5
+ "origin_strat_2",
6
+ "origin_strat_3",
7
+ "umar_pads",
8
+ "umar_pads_500"
9
+ ],
10
+ "carve_salt": "agents-v3",
11
+ "carve_mode": "origin-stratified",
12
+ "excluded": {
13
+ "duplicate_text": 253,
14
+ "eval_text": 77
15
+ },
16
+ "eval_docs_checked": 128
17
+ },
18
+ "splits": {
19
+ "train": {
20
+ "rows": 16843,
21
+ "spans": 12891,
22
+ "labels": {
23
+ "NAMED_DATA": 9723,
24
+ "DESCRIPTIVE_DATA": 3142,
25
+ "VAGUE_DATA": 26
26
+ },
27
+ "documents": 10265,
28
+ "origins": {
29
+ "fcv_pads_east_africa": 392,
30
+ "jdc_operational": 11,
31
+ "prwp": 5351,
32
+ "reliefweb": 885,
33
+ "umar_pads": 10204
34
+ },
35
+ "origin_documents": {
36
+ "fcv_pads_east_africa": 357,
37
+ "jdc_operational": 4,
38
+ "prwp": 5327,
39
+ "reliefweb": 880,
40
+ "umar_pads": 3697
41
+ }
42
+ },
43
+ "val": {
44
+ "rows": 2122,
45
+ "spans": 1665,
46
+ "labels": {
47
+ "NAMED_DATA": 1212,
48
+ "DESCRIPTIVE_DATA": 451,
49
+ "VAGUE_DATA": 2
50
+ },
51
+ "documents": 1283,
52
+ "origins": {
53
+ "fcv_pads_east_africa": 47,
54
+ "prwp": 667,
55
+ "reliefweb": 111,
56
+ "umar_pads": 1297
57
+ },
58
+ "origin_documents": {
59
+ "fcv_pads_east_africa": 45,
60
+ "prwp": 666,
61
+ "reliefweb": 110,
62
+ "umar_pads": 462
63
+ }
64
+ },
65
+ "holdout": {
66
+ "rows": 2108,
67
+ "spans": 1611,
68
+ "labels": {
69
+ "DESCRIPTIVE_DATA": 383,
70
+ "NAMED_DATA": 1225,
71
+ "VAGUE_DATA": 3
72
+ },
73
+ "documents": 1283,
74
+ "origins": {
75
+ "fcv_pads_east_africa": 49,
76
+ "prwp": 670,
77
+ "reliefweb": 112,
78
+ "umar_pads": 1277
79
+ },
80
+ "origin_documents": {
81
+ "fcv_pads_east_africa": 45,
82
+ "prwp": 666,
83
+ "reliefweb": 110,
84
+ "umar_pads": 462
85
+ }
86
+ }
87
+ }
88
+ }