Dataset Viewer
Auto-converted to Parquet Duplicate
concept
stringlengths
4
8
greek_strong
stringlengths
1
4
greek_lemma
stringlengths
0
14
source_strong
stringlengths
1
64
count
int32
1
2.69k
share_in_concept
float64
0
1
share_in_greek
float64
0
1
n_jev
int32
0
2.4k
n_opus
int32
0
375
n_opus_jev
int32
0
654
jev_confidence_mean
float64
0
1
H1.1
3962
πατήρ
3962
485
0.9759
0.4659
433
13
39
0.8977
H1.1
3588
ὁ
3588
9
0.0181
0.0062
9
0
0
0.9233
H1.1
2228
ἤ
2228
1
0.002
0.0833
1
0
0
0.99
H1.1
3146
μαστιγόω
3146
1
0.002
0.3333
1
0
0
0.98
H1.1
3384
μήτηρ
3384
1
0.002
0.0055
1
0
0
1
H1.10
3962
πατήρ
3962
1
1
0.001
1
0
0
1
H1.11
3962
πατήρ
3962
9
1
0.0086
3
4
2
0.5633
H1.12
3962
πατήρ
3962
7
0.875
0.0067
7
0
0
0.88
H1.12
2316
θεός
2316
1
0.125
0.0004
1
0
0
0.98
H1.2
3962
πατήρ
3962
8
1
0.0077
0
8
0
0.6887
H1.3
3962
πατήρ
3962
274
0.9928
0.2632
199
13
62
0.8065
H1.3
2316
θεός
2316
1
0.0036
0.0004
0
0
1
0.67
H1.3
3588
ὁ
3588
1
0.0036
0.0007
1
0
0
0.92
H1.4
3962
πατήρ
3962
57
1
0.0548
46
5
6
0.7732
H1.5
3962
πατήρ
3962
54
0.9643
0.0519
36
7
11
0.753
H1.5
3588
ὁ
3588
1
0.0179
0.0007
1
0
0
0.77
H1.5
3965
πατριά
3965
1
0.0179
0.05
0
1
0
0.62
H1.6
3962
πατήρ
3962
55
0.8462
0.0528
42
3
10
0.7667
H1.6
3965
πατριά
3965
8
0.1231
0.4
7
0
1
0.7937
H1.6
3967
πατρικός
3967
1
0.0154
1
0
1
0
0.63
H1.6
4771
σύ
5216
1
0.0154
0.0222
1
0
0
0.89
H1.7
3962
πατήρ
3962
41
0.9318
0.0394
40
0
1
0.9629
H1.7
3965
πατριά
3965
2
0.0455
0.1
2
0
0
0.96
H1.7
3966
πατριάρχης
3966
1
0.0227
0.5
1
0
0
0.97
H1.8
3962
πατήρ
3962
5
0.625
0.0048
5
0
0
0.878
H1.8
3965
πατριά
3965
2
0.25
0.1
2
0
0
0.91
H1.8
2596
κατά
2596
1
0.125
0.0263
0
0
1
0.67
H1.9
3962
πατήρ
3962
38
0.95
0.0365
30
8
0
0.8708
H1.9
2730
κατοικέω
2730
1
0.025
0.0028
1
0
0
0.92
H1.9
5207
υἱός
5207
1
0.025
0.0003
1
0
0
0.93
H6.1
622
ἀπόλλυμι
622
45
0.9375
0.2406
37
1
7
0.8256
H6.1
684
ἀπώλεια
684
2
0.0417
0.0714
1
0
1
0.695
H6.1
1726
ἐναντίον
1726
1
0.0208
0.0085
1
0
0
0.71
H6.2
622
ἀπόλλυμι
622
6
1
0.0321
2
1
3
0.615
H6.3
622
ἀπόλλυμι
622
24
0.8571
0.1283
18
0
6
0.7992
H6.3
623
Ἀπολλύων
623
1
0.0357
1
1
0
0
0.71
H6.3
684
ἀπώλεια
684
1
0.0357
0.0357
1
0
0
0.91
H6.3
1287
διασκορπίζω
1287
1
0.0357
0.0526
1
0
0
0.77
H6.3
1821
ἐξαποστέλλω
1821
1
0.0357
0.0063
1
0
0
0.86
H6.4
622
ἀπόλλυμι
622
3
1
0.016
2
1
0
0.7967
H6.5
622
ἀπόλλυμι
622
6
1
0.0321
5
0
1
0.9183
H6.6
622
ἀπόλλυμι
622
13
0.8667
0.0695
9
2
2
0.8154
H6.6
846
αὐτός
848
1
0.0667
0.0077
0
1
0
0.51
H6.6
1587
ἐκλείπω
1587
1
0.0667
0.0145
1
0
0
0.98
H6.7
622
ἀπόλλυμι
622
6
0.8571
0.0321
5
0
1
0.9217
H6.7
1842
ἐξολεθρεύω
1842
1
0.1429
0.0103
1
0
0
0.94
H6.8
622
ἀπόλλυμι
622
5
0.8333
0.0267
5
0
0
0.958
H6.8
577
ἀποβάλλω
577
1
0.1667
0.25
0
0
1
0.48
H6.9
622
ἀπόλλυμι
622
6
1
0.0321
2
3
1
0.56
H14.1
2309
θέλω
2309
17
0.7083
0.3208
11
0
6
0.7359
H14.1
1014
βούλομαι
1014
6
0.25
0.2143
5
0
1
0.8233
H14.1
2106
εὐδοκέω
2106
1
0.0417
0.0909
0
0
1
0.51
H14.2
1014
βούλομαι
1014
1
0.3333
0.0357
0
1
0
0.34
H14.2
2106
εὐδοκέω
2106
1
0.3333
0.0909
0
1
0
0.57
H14.2
2309
θέλω
2309
1
0.3333
0.0189
0
1
0
0.7
H14.3
2309
θέλω
2309
2
0.6667
0.0377
1
1
0
0.735
H14.3
1014
βούλομαι
1014
1
0.3333
0.0357
0
1
0
0.46
H14.5
544
ἀπειθέω
544
1
0.3333
0.05
1
0
0
0.99
H14.5
1262
διαλύω
1262
1
0.3333
0.3333
1
0
0
0.96
H14.5
2309
θέλω
2309
1
0.3333
0.0189
1
0
0
1
H34.1
3993
πένης
3993
8
0.4
0.1633
8
0
0
0.8425
H34.1
4434
πτωχός
4434
6
0.3
0.1071
2
0
4
0.6067
H34.1
102
ἀδύνατος
102
2
0.1
0.2857
0
0
2
0.58
H34.1
772
ἀσθενής
772
1
0.05
0.1429
0
0
1
0.52
H34.1
1729
ἐνδεής
1729
1
0.05
1
1
0
0
0.92
H34.1
5013
ταπεινόω
5013
1
0.05
0.0294
1
0
0
0.72
H34.1
5503
χῆρος
5503
1
0.05
0.0227
1
0
0
0.8
H34.2
3993
πένης
3993
16
0.8
0.3265
4
8
4
0.4338
H34.2
4434
πτωχός
4434
2
0.1
0.0357
0
1
1
0.36
H34.2
120
ἀθυμέω
120
1
0.05
1
0
1
0
0.06
H34.2
560
ἀπελπίζω
560
1
0.05
1
0
0
1
0
H56.5
684
ἀπώλεια
684
1
1
0.0357
1
0
0
0.88
H60.1
3997
πένθος
3997
5
0.8333
1
4
1
0
0.84
H60.1
2805
κλαυθμός
2805
1
0.1667
0.0556
1
0
0
0.88
H68.1
3037
λίθος
3037
39
1
0.2021
36
2
1
0.8838
H68.10
3037
λίθος
3037
5
1
0.0259
4
1
0
0.81
H68.12
3037
λίθος
3037
37
0.925
0.1917
36
0
1
0.9332
H68.12
3588
ὁ
3588
2
0.05
0.0014
2
0
0
0.995
H68.12
3586
ξύλον
3586
1
0.025
0.0048
1
0
0
0.97
H68.13
3037
λίθος
3037
9
0.6923
0.0466
9
0
0
0.9478
H68.13
3035
λίθινος
3035
3
0.2308
1
2
1
0
0.89
H68.13
3586
ξύλον
3586
1
0.0769
0.0048
1
0
0
0.89
H68.14
3037
λίθος
3037
1
1
0.0052
1
0
0
1
H68.17
3037
λίθος
3037
1
1
0.0052
1
0
0
1
H68.18
3037
λίθος
3037
3
0.75
0.0155
3
0
0
0.97
H68.18
2532
καί
2532
1
0.25
0.0023
1
0
0
0.98
H68.2
3037
λίθος
3037
18
1
0.0933
18
0
0
0.8767
H68.3
3037
λίθος
3037
10
0.9091
0.0518
8
2
0
0.793
H68.3
3588
ὁ
3588
1
0.0909
0.0007
1
0
0
0.75
H68.4
3037
λίθος
3037
12
1
0.0622
7
3
2
0.6908
H68.5
3037
λίθος
3037
2
1
0.0104
0
2
0
0.325
H68.6
3037
λίθος
3037
9
1
0.0466
5
0
4
0.7478
H68.7
3037
λίθος
3037
3
0.75
0.0155
2
0
1
0.7167
H68.7
3588
ὁ
3588
1
0.25
0.0007
0
0
1
0.35
H68.8
3037
λίθος
3037
16
1
0.0829
16
0
0
0.9862
H68.9
3037
λίθος
3037
21
0.9545
0.1088
16
2
3
0.7595
H68.9
2532
καί
2532
1
0.0455
0.0023
1
0
0
0.72
H113.1
2962
κύριος
2962
71
0.8256
0.2046
14
42
15
0.4534
H113.1
3588
ὁ
3588
15
0.1744
0.0103
4
9
2
0.5153
H113.2
2962
κύριος
2962
13
0.8667
0.0375
2
7
4
0.4331
End of preview. Expand in Data Studio

Hebrew word concepts, model-labelled (v0.6)

This dataset is model output. No person has checked it. For contested Hebrew content words (where translations split a word's renderings, or one UBS sense covers renderings that differ), each word gets a set of concepts, finer than or equal to its UBS senses, and each occurrence its concept: 154,598 occurrences of 1,234 words, 8,555 concepts, labelled in 6 stages (genesis, torah, narrative, poetry_prophets, b1, b2). Labellers over all: jev 112,248, opus 17,086, opus+jev 23,177, unresolved 2,087.

How it was made

  • Concepts per word: Claude Opus 5.5 (claude-opus-5-5) read up to 80 occurrences of the word across the Tanakh: the Hebrew clause (MACULA lemma and morphology), the word's UBS senses and their definitions, the Septuagint lemma at each place, and how 28 openly licensed translations group the occurrences (cluster ids and counts only, no translation text). It did not see MACULA's English gloss, any translation's wording or BHSA data. It proposed codes H<strong>.<n> with a one-line definition each. A word defined in an earlier stage keeps its concepts and codes. Stage b1 defines no concepts (it labels more occurrences of words already defined); stage b2 defines concepts for 20 words that have no UBS sense, without UBS senses (see its section).
  • Concept per occurrence: Jev (TypeSafe) and Claude Opus 5.5 chose among the word's concepts; each stage's rule and its labeller per row are below and in tokens (labeller, each run's label, Jev's confidence).
  • Run by make-tanakh-research to bcv-query's specification.

Translations whose grouping was shown (public domain or CC BY-SA 4.0, verified): arb_vdv, cek_akb, ces_bkr, cmn_cu1, cth_bcf, deu_elo, epo_bfb, fra_jnd, hat_pds, haw_868, hlt_mcs, hne_twf, isl_bib, ita_dio, kor_old, kos_pub, lat_clv, mya_jvb, nay_afa, nld_, pes_pbs, ron_924, rus_syn, spa_r09, srp_865, ton_png, vie_1934, yom_bib. 39 other translations were held back for the independent check.

How good it is, per stage

Genesis: genesis (8,308 occurrences)

  • Scope: Genesis, 341 contested non-name lexemes, 8,308 tokens; concepts defined in step A from up to 80 occurrences per lexeme across the whole Tanakh; the closed chapters (EST, DAN, JON, GEN 22/37/44) left out of both steps (0 closed keys in tokens.tsv).
  • Labelling: Each token's concept is Jev's when both Jev runs (concept list in two orders) agree at confidence >= 0.8; else Opus's when both Opus runs (two orders) agree; else unresolved. The labeller is named per row (tokens.tsv: labeller, jev_run1, jev_run2, jev_confidence, opus_run1, opus_run2). Jev version: jev-latest = jev-1.14.0. Labellers: jev 5,516, opus 2,629, unresolved 163.
  • Two labellings of the same occurrences (model with itself, not human agreement): Jev run 1 = run 2 (all tokens): 0.919; Opus run 1 = run 2 (the 2,791 tokens below the band): 0.942; Jev run 1 = the Opus agreed label on those tokens: 0.679. 9.3 concepts per word against 2.3 UBS senses.
  • Independent check on 39 translations never shown: the concept predicts how a translation groups a word's occurrences better than the UBS sense. NMI 0.368 against 0.204, +0.164 [+0.159, +0.169]; AMI +0.077 [+0.073, +0.081]; purity +0.062 [+0.059, +0.065] (95% intervals, 7,104 word-translation pairs).
  • Use in drafting (Indonesian word list keyed by (word, concept) against (word, UBS sense), learned and scored on separate halves): on the 2,201 occurrences where the two differ, 0.176 -> 0.368, +19.2 [+16.4, +22.0] points; over 1,532 verses word F1 44.20 -> 45.55, chrF1 45.63 -> 46.65 (point figures).
  • Cost (list price): USD 42.83 (step A (Opus) 28.99; step B Jev, two runs 0.80; step B Opus below the band, two runs 13.04).

Torah: EXO, LEV, NUM, DEU (27,230 occurrences)

  • Scope: EXO, LEV, NUM, DEU: 641 contested non-name lexemes (381 defined in this stage, 260 reused from earlier stages), 27230 tokens; contested by Genesis's test with each book's own track D fold; concepts defined from up to 80 occurrences across the Tanakh; the closed chapters (EST, DAN, JON, GEN 22/37/44) left out.
  • Labelling: Each token's concept is Jev's when its run 1 has confidence >= 0.9 (no run 2 asked) or both Jev runs (concept list in two orders) agree at confidence >= 0.7; else Opus's when Opus's single run agrees with Jev's run 1 (labeller opus+jev); else Opus's when both Opus runs agree (labeller opus); else unresolved. The labeller is named per row. Jev version: jev-latest = jev-1.14.0. Labellers: jev 19,584, opus 2,655, opus+jev 4,701, unresolved 290.
  • Two labellings of the same occurrences (model with itself, not human agreement): Jev run 1 = run 2 (where run 2 was asked): 0.824; Opus run 1 = run 2 (where run 2 was asked: Opus run 1 differs from Jev): 0.901. 8.1 concepts per word against 2.4 UBS senses.
  • Independent check on 39 translations never shown: the concept predicts how a translation groups a word's occurrences better than the UBS sense. NMI 0.343 against 0.200, +0.143 [+0.137, +0.148]; AMI +0.079 [+0.073, +0.084]; purity +0.055 [+0.051, +0.058] (95% intervals, 7,609 word-translation pairs).
  • Use in drafting (Indonesian word list keyed by (word, concept) against (word, UBS sense), learned and scored on separate halves): on the 6,330 occurrences where the two differ, 0.193 -> 0.389, +19.6 [+17.6, +21.4] points; over 4,288 verses word F1 43.22 -> 44.16, chrF1 44.08 -> 45.02 (point figures).
  • Cost (list price): USD 51.45 (step A (Opus, new lexemes) 26.54; step B Jev 1.97; step B Opus below the band 22.94).

Narrative: JOS, JDG, RUT, 1SA, 2SA, 1KI, 2KI, 1CH, 2CH, EZR, NEH (42,881 occurrences)

  • Scope: JOS, JDG, RUT, 1SA, 2SA, 1KI, 2KI, 1CH, 2CH, EZR, NEH: 681 contested non-name lexemes (200 defined in this stage, 481 reused from earlier stages), 42881 tokens; contested by Genesis's test with each book's own track D fold; concepts defined from up to 80 occurrences across the Tanakh; the closed chapters (EST, DAN, JON, GEN 22/37/44) left out.
  • Labelling: Each token's concept is Jev's when its run 1 has confidence >= 0.9 (no run 2 asked) or both Jev runs (concept list in two orders) agree at confidence >= 0.7; else Opus's when Opus's single run agrees with Jev's run 1 (labeller opus+jev); else Opus's when both Opus runs agree (labeller opus); else unresolved. The labeller is named per row. Jev version: jev-latest = jev-1.14.0. Labellers: jev 29,114, opus 4,993, opus+jev 8,095, unresolved 679.
  • Two labellings of the same occurrences (model with itself, not human agreement): Jev run 1 = run 2 (where run 2 was asked): 0.839; Opus run 1 = run 2 (where run 2 was asked: Opus run 1 differs from Jev): 0.881. 8.1 concepts per word against 2.9 UBS senses.
  • Independent check on 39 translations never shown: the concept predicts how a translation groups a word's occurrences better than the UBS sense. NMI 0.325 against 0.193, +0.132 [+0.127, +0.138]; AMI +0.074 [+0.069, +0.079]; purity +0.050 [+0.047, +0.053] (95% intervals, 7,376 word-translation pairs).
  • Use in drafting (Indonesian word list keyed by (word, concept) against (word, UBS sense), learned and scored on separate halves): on the 9,137 occurrences where the two differ, 0.208 -> 0.364, +15.6 [+14.2, +17.1] points; over 6,809 verses word F1 46.44 -> 47.42, chrF1 46.68 -> 47.59 (point figures).
  • Cost (list price): USD 52.37 (step A (Opus, new lexemes) 11.19; step B Jev 3.35; step B Opus below the band 37.83).

Poetry and prophets: PSA, JOB, PRO, ECC, SNG, LAM, ISA, JER, EZK, HOS, JOL, AMO, OBA, MIC, NAM, HAB, ZEP, HAG, ZEC, MAL (50,961 occurrences)

  • Scope: PSA, JOB, PRO, ECC, SNG, LAM, ISA, JER, EZK, HOS, JOL, AMO, OBA, MIC, NAM, HAB, ZEP, HAG, ZEC, MAL: 994 contested non-name lexemes (292 defined in this stage, 702 reused from earlier stages), 50961 tokens; contested by Genesis's test with each book's own track D fold; concepts defined from up to 80 occurrences across the Tanakh; the closed chapters (EST, DAN, JON, GEN 22/37/44) left out.
  • Labelling: Each token's concept is Jev's when its run 1 has confidence >= 0.9 (no run 2 asked) or both Jev runs (concept list in two orders) agree at confidence >= 0.7; else Opus's when Opus's single run agrees with Jev's run 1 (labeller opus+jev); else Opus's when both Opus runs agree (labeller opus); else unresolved. The labeller is named per row. Jev version: jev-latest = jev-1.14.0. Labellers: jev 32,816, opus 6,809, opus+jev 10,381, unresolved 955.
  • Two labellings of the same occurrences (model with itself, not human agreement): Jev run 1 = run 2 (where run 2 was asked): 0.809; Opus run 1 = run 2 (where run 2 was asked: Opus run 1 differs from Jev): 0.877. 7.6 concepts per word against 3.4 UBS senses.
  • Independent check on 39 translations never shown: the concept predicts how a translation groups a word's occurrences better than the UBS sense. NMI 0.366 against 0.224, +0.142 [+0.138, +0.146]; AMI +0.061 [+0.057, +0.064]; purity +0.055 [+0.053, +0.058] (95% intervals, 14,391 word-translation pairs).
  • Use in drafting (Indonesian word list keyed by (word, concept) against (word, UBS sense), learned and scored on separate halves): on the 12,296 occurrences where the two differ, 0.201 -> 0.336, +13.5 [+12.2, +14.8] points; over 9,815 verses word F1 41.51 -> 42.50, chrF1 43.23 -> 44.07 (point figures).
  • Cost (list price): USD 67.25 (step A (Opus, new lexemes) 13.40; step B Jev 3.95; step B Opus below the band 49.90).

B1: b1 (11,827 occurrences)

  • Scope: MACULA Hebrew noun/verb/adj parts, not proper (MACULA type), closed chapters out, key not in hebrew-word-concepts-model v0.5 tokens.parquet, lexeme with concept rows in v0.5.
  • Kept: Jev run 1 confidence >= 0.8 and, where run 2 was asked, the same concept; 11,827 of 19,675 occurrences asked. The rest are not in tokens.
  • Check on translations never shown, against the UBS sense: AMI 0.172 against 0.112, difference +0.060 [+0.056, +0.064] (95% interval, 11,569 word-translation pairs).
  • English: Jev read the concepts' English definitions this once, by the user's decision of 2026-11-17; later designs are to work without English descriptions.
  • Cost (list price): Jev USD 1.32.

B2: b2 (13,391 occurrences)

  • Scope: 20 lexemes with no UBS sense on any occurrence (>= 20 tokens), closed chapters out.
  • Kept: both Jev runs give the same concept (run 2 lists the concepts in reverse order); run 1 alone where its confidence >= 0.9 (run 2 not asked); 13,391 of 15,329 occurrences asked. The rest are not in tokens.
  • Check on translations never shown, against one class per word: AMI 0.081 against 0.000, difference +0.081 [+0.073, +0.089] (95% interval, 513 word-translation pairs).
  • Check on translations never shown, against the Septuagint lemma: AMI 0.081 against 0.035, difference +0.046 [+0.041, +0.051] (95% interval, 513 word-translation pairs).
  • Two Jev runs agree on 83.9% of the 12,026 occurrences asked twice (model with itself); the bar set beforehand was 90%, so only agreeing runs are kept.
  • Concepts for these words were defined without UBS senses (none exist for them): from the Septuagint lemma, how the shown translations group the occurrences, and the Hebrew clause. Concept codes have ubs_sense = none.
  • English: Jev read the concepts' English definitions this once, by the user's decision of 2026-11-17; later designs are to work without English descriptions.
  • Cost (list price): Jev USD 1.07.

Files

file content
data/tokens.parquet stage, key (MACULA Hebrew word id), lang (hbo, or arc for Aramaic), concept, labeller, and each labelling run (labels, Jev's confidence)
data/concepts.parquet code, strong, ubs_sense (the UBS sense the concept refines)
data/concept_greek.parquet per concept, the Greek lexemes of the Septuagint at its occurrences (below)
definitions.tsv the model's one-line English definition per concept, kept apart from the data; examples quoting a phrase were removed

concept_greek: the concepts seen through the Septuagint

109,192 of the 152,511 labelled occurrences have a Greek word in MACULA Hebrew's Septuagint alignment. Counted per concept, that gives 13,148 concept-Greek pairs over 6,075 concepts and 1,953 Greek lexemes. The pairs are counts over model labels; the alignment is MACULA's.

column description
concept the concept code
greek_strong the Greek lexeme's Strong's number as MACULA's Greek NT numbers it, one per lemma (plain digits)
greek_lemma that lexeme's lemma in MACULA's Greek NT; empty for the 16 numbers the NT does not have
source_strong the number or numbers the alignment itself gives (it numbers some inflected forms apart, e.g. 2036 for εἶπεν; mapped to the NT's number by bcv-query's resources/lxx_bridge_tiers/greek_numbers.tsv)
count occurrences of the concept rendered by this Greek lexeme
share_in_concept count over the concept's occurrences that have a Greek word
share_in_greek count over all labelled occurrences rendered by this Greek lexeme; low when the Greek word serves several concepts
n_jev, n_opus, n_opus_jev how many of the count occurrences each labeller labelled (labeller in tokens)
jev_confidence_mean mean of Jev's confidence over the occurrences that have one

Only occurrences in this dataset are counted: contested words outside the closed books, so the shares are not the whole Septuagint's.

Occurrences in Genesis 22, 37 and 44, Esther, Daniel and Jonah are left out (held back for evaluation elsewhere).

230 occurrences are Aramaic (Ezra's Aramaic chapters, lang arc). Their concepts are keyed by the Aramaic lexeme's own Strong's number (the codes do not collide with Hebrew ones); bcv-commons/aramaic-hebrew-equivalents links those lexemes to Hebrew.

Licence

CC BY-SA 4.0. Inputs: MACULA Hebrew (CC BY 4.0), the UBS dictionaries' senses (CC BY-SA 4.0), GLAUx Septuagint (CC BY-SA 4.0), and groupings derived from the translations listed above (public domain, CC BY-SA 4.0). Credit: Claude Opus 5.5 (Anthropic) and Jev (TypeSafe) as labellers, make-tanakh-research (the runs and measurements), MACULA (Clear Bible / Biblica), United Bible Societies, GLAUx (KU Leuven), the translation teams of the editions above, and this dataset.

Downloads last month
126