Datasets:
Tasks:
Feature Extraction
Formats:
parquet
Languages:
Multiple languages
Size:
100K - 1M
ArXiv:
License:
File size: 55,727 Bytes
f326e95 3055cc8 cba561e 9206e91 76d8a43 b92c43f 9206e91 4352f73 9206e91 857fc54 ee7a22f 3055cc8 b158bbc 3055cc8 b158bbc 76d8a43 3055cc8 58b87db 3055cc8 58b87db 3055cc8 58b87db f326e95 58b87db 3055cc8 58b87db 3055cc8 58b87db f326e95 58b87db f326e95 58b87db 3055cc8 58b87db 3055cc8 58b87db 3055cc8 58b87db 3055cc8 58b87db 3055cc8 58b87db 3055cc8 58b87db 3055cc8 58b87db 3055cc8 9206e91 76d8a43 cba561e 76d8a43 cba561e 76d8a43 cba561e 76d8a43 cba561e 76d8a43 b92c43f cba561e b92c43f cba561e b92c43f cba561e b92c43f cba561e b92c43f cba561e b92c43f cba561e b92c43f cba561e b92c43f cba561e b92c43f cba561e b92c43f cba561e b92c43f cba561e b92c43f cba561e b92c43f cba561e db9df30 cba561e db9df30 f326e95 db9df30 cba561e db9df30 cba561e db9df30 4352f73 db9df30 4352f73 db9df30 364a0dd b158bbc 364a0dd b158bbc 364a0dd b158bbc 364a0dd b158bbc 364a0dd 857fc54 8f78213 857fc54 cba561e 857fc54 8f78213 857fc54 ee7a22f f326e95 ee7a22f 8f78213 85ed145 cba561e 85ed145 3055cc8 ee7a22f cba561e ee7a22f 4352f73 ee7a22f 8f78213 85ed145 3055cc8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 | # ruff: noqa: T201 # CLI build-script report payload (CLAUDE.md whitelist)
"""Build NuBerea/concept-analysis — consolidated analytical layer for concept topology.
Absorbs all of NuBerea/concept-trajectories and migrates three analytical configs
from NuBerea/concept-graph (community_candidate, community_membership, concept_centroids).
Boundary principle:
concept-graph = topology (span_node, span_edge, concept_node, concept_span_edge, ...)
concept-analysis = derived outputs (centroids, communities, trajectories, drift)
Output configs (20 total, writing 22 parquet dirs — pseu builds write deut+pseu each):
concept_verse_edge — 28,909 rows; concept×verse, LXX+NT strata, shlm-grc-en 768-dim
concept_trajectory — 112 rows; concept×stratum centroids, LXX+NT
concept_verse_edge_sp — 59,478 rows; concept×verse, LXX+NT+Vulgate, SPhilBerta 768-dim
concept_trajectory_sp — 190 rows; concept×stratum centroids, LXX+NT+Vulgate
concept_verse_edge_bt — 59,478 rows; concept_verse_edge_sp + bt_layer + plate_segment (T8)
concept_trajectory_bt — ~960 rows; concept×bt_layer centroids (T8, match_grain=pericope)
concept_verse_edge_deut — ~5,857 rows; Deuterocanon verses, nearest-centroid assignment
concept_verse_edge_pseu — ~15,043 rows; Pseudepigrapha verses, nearest-centroid assignment
concept_trajectory_deut — per-concept centroids for Deuterocanon stratum
concept_trajectory_pseu — per-concept centroids for Pseudepigrapha stratum
concept_drift — 78 rows; Schlattmann&Vogl 2024 drift metrics per concept×transition
concept_drift_itp — 78 rows; PSEU-extended drift (5 pairs: lxx→deut/pseu, deut/pseu→nt, deut→pseu)
concept_drift_hamilton — 5,023 rows; Hamilton/CADE word2vec drift per lemma (3 languages)
community_candidate — 55 rows; community summaries (from concept-graph)
community_membership — 2,412 rows; pericope→community assignments (from concept-graph)
concept_centroids — 156 rows; L1+L2 bi_encoder_v1 384-dim centroids (from concept-graph)
concept_aliases — 567 rows; TF-IDF English term → bt1_xxx concept_id list mappings
concept_verse_edge_patr — ~40K rows; patristic GRC segments, nearest-centroid assignment
concept_verse_edge_philo — ~7.9K rows; Philo of Alexandria GRC segments, nearest-centroid
concept_verse_edge_josephus — ~12.7K rows; Josephus GRC segments, nearest-centroid
concept_verse_edge_targum — ~28K rows; Targum English translations, nearest-centroid
concept_verse_edge_talmud — ~200K rows; Mishnah+Tosefta English translations, nearest-centroid
Source data (local build dirs + HF for HF-native configs):
huggingface/candidates/build/concept-trajectories/data/
huggingface/candidates/build/concept-graph/data/
huggingface/candidates/build/concept-trajectories/data/concept_drift_analysis.csv
concept_aliases: HF-native — NuBerea/AI-models bi_encoder_v1 + question_bin_assignment.json
+ NuBerea/concept-graph concept_node centroids
concept_verse_edge_bt / concept_trajectory_bt: T8 — read NuBerea/pericope-genealogy
composition_dag (local until genealogy HF upload); promote concept-analysis to T8 for
bt-indexed configs only.
Usage:
python scripts/build.py # build all configs
python scripts/build.py --config X # build only config X
For HF-native reproducibility of the trajectory configs, see verify.py in this directory.
"""
from __future__ import annotations
import argparse
import pathlib
import sys
import pandas as pd
BUILD = pathlib.Path("huggingface/candidates/build")
OUT = BUILD / "concept-analysis/data"
# Make hf_registry importable
_THIS = pathlib.Path(__file__).resolve()
_CANDIDATES = _THIS.parents[2]
if str(_CANDIDATES) not in sys.path:
sys.path.insert(0, str(_CANDIDATES))
from hf_registry import load_hf_or_local
def _parquet_out(config: str) -> pathlib.Path:
p = OUT / config
p.mkdir(parents=True, exist_ok=True)
return p / "train-00000-of-00001.parquet"
def import_trajectory_config(name: str) -> None:
"""Import a trajectory config from NuBerea/concept-trajectories (HF-native, local fallback)."""
ds = load_hf_or_local("NuBerea/concept-analysis", name)
df = ds.to_pandas()
dst = _parquet_out(name)
df.to_parquet(dst, index=False)
print(f" {name}: {len(df):,} rows, {len(df.columns)} cols — imported from NuBerea/concept-trajectories")
def build_concept_drift() -> None:
"""Compute drift inline from concept_trajectory_sp + concept_verse_edge_sp (HF-native)."""
sys.path.insert(0, str(_THIS.parent))
from analyze_concept_drift import compute_drift_dataframe
traj = load_hf_or_local("NuBerea/concept-analysis", "concept_trajectory_sp").to_pandas()
cve = load_hf_or_local("NuBerea/concept-analysis", "concept_verse_edge_sp").to_pandas()
df = compute_drift_dataframe(traj, cve)
dst = _parquet_out("concept_drift")
df.to_parquet(dst, index=False)
print(f" concept_drift: {len(df)} rows, {len(df.columns)} cols — computed inline (HF-native)")
def import_community_config(name: str) -> None:
"""Import a community config from NuBerea/concept-graph (HF-native, local fallback)."""
ds = load_hf_or_local("NuBerea/concept-graph", name)
df = ds.to_pandas()
dst = _parquet_out(name)
df.to_parquet(dst, index=False)
print(f" {name}: {len(df):,} rows, {len(df.columns)} cols — imported from NuBerea/concept-graph")
def build_concept_centroids() -> None:
"""Build concept_centroids from NuBerea/concept-graph concept_node (HF-native).
The concept_node config carries each L1/L2 BERTopic node's centroid inline as a
`centroid` column (bi_encoder_v1 384-dim vectors). We project to a 3-column
parquet (concept_id, level, centroid) matching the historical schema.
"""
ds = load_hf_or_local("NuBerea/concept-graph", "concept_node")
df = ds.to_pandas()
df = df[df["level"].isin([1, 2])].copy()
df["level"] = df["level"].map({1: "l1", 2: "l2"})
df = df[["concept_id", "level", "centroid"]].sort_values(["level", "concept_id"]).reset_index(drop=True)
dst = _parquet_out("concept_centroids")
df.to_parquet(dst, index=False)
per_level = df.groupby("level").size().to_dict()
print(f" concept_centroids: {len(df)} rows ({per_level}) — built from concept_node (HF-native)")
def build_concept_drift_hamilton() -> None:
"""Run the Hamilton/CADE word2vec drift analysis (requires gensim).
Delegates to build_hamilton.py, which trains word2vec per corpus slice,
aligns via orthogonal Procrustes, and computes drift_score per shared lemma.
Three independent analyses: Hebrew EBH→LBH, Greek LXX→NT, Latin VG_OT→VG_NT.
Requires Python 3.12 venv with gensim installed:
/tmp/gensim_venv/bin/python scripts/build_hamilton.py
Or equivalently:
python scripts/build.py --config concept_drift_hamilton
(which calls build_hamilton.py via subprocess using the same interpreter that
has gensim, or prints existing stats if the parquet already exists.)
"""
import subprocess
import sys
parquet = OUT / "concept_drift_hamilton" / "train-00000-of-00001.parquet"
# Try running build_hamilton.py with current interpreter first; fall back to gensim venv
interpreters = [sys.executable, "/tmp/gensim_venv/bin/python"]
for interp in interpreters:
try:
result = subprocess.run(
[interp, "scripts/build_hamilton.py"],
check=True, capture_output=False,
)
break
except (subprocess.CalledProcessError, FileNotFoundError):
continue
else:
if parquet.exists():
df = pd.read_parquet(parquet)
print(f" concept_drift_hamilton: {len(df):,} rows (pre-built; gensim unavailable)")
else:
raise RuntimeError(
"build_hamilton.py failed and no pre-built parquet found. "
"Run: /tmp/gensim_venv/bin/python scripts/build_hamilton.py"
)
return
df = pd.read_parquet(parquet)
print(f" concept_drift_hamilton: {len(df):,} rows — Hamilton/CADE word2vec drift")
def _build_verse_bt_lookup(dag: "pd.DataFrame", nrsv: "pd.DataFrame") -> dict:
"""Return dict mapping OSIS verse ref (dot format) → (bt_layer, plate_segment).
Covers three node types in composition_dag:
- otp:/pseu: rows with osis_start/osis_end → range-expanded to constituent verses
- single-verse rows (unit_id matches 'Book.ch.v') → direct mapping
"""
import re
# Parse OSIS ref to (chapter, verse) numeric key for range comparison
_osis_key_re = re.compile(r'^[^.]+\.(\d+)\.(\d+)$')
def _key(ref: str) -> tuple[int, int] | None:
m = _osis_key_re.match(ref)
if not m:
return None
return (int(m.group(1)), int(m.group(2)))
verse_bt: dict = {}
# ── 1. Range-based: otp + pseu rows (have osis_start / osis_end) ────────
range_rows = dag[
dag['unit_id'].str.startswith(('otp:', 'pseu:'), na=False)
& dag['osis_start'].ne('')
][['osis_start', 'osis_end', 'bt_layer', 'plate_segment']].copy()
range_rows['_book'] = range_rows['osis_start'].str.split('.').str[0]
range_rows['_sk'] = range_rows['osis_start'].apply(_key)
range_rows['_ek'] = range_rows['osis_end'].apply(_key)
range_rows = range_rows.dropna(subset=['_sk', '_ek'])
# Build a numeric sort key for merge_asof (ch * 10000 + v fits all Bible ranges)
range_rows['_start_num'] = range_rows['_sk'].apply(lambda k: k[0] * 10000 + k[1])
range_rows['_end_num'] = range_rows['_ek'].apply(lambda k: k[0] * 10000 + k[1])
nrsv_ot = nrsv[nrsv['osis_ref'].str.contains(r'^\w+\.\d+\.\d+$', regex=True, na=False)].copy()
nrsv_ot['_book'] = nrsv_ot['osis_ref'].str.split('.').str[0]
nrsv_ot['_num'] = nrsv_ot['osis_ref'].apply(
lambda r: (lambda k: k[0] * 10000 + k[1] if k else None)(_key(r))
)
nrsv_ot = nrsv_ot.dropna(subset=['_num'])
nrsv_ot['_num'] = nrsv_ot['_num'].astype(int)
for book, book_perps in range_rows.groupby('_book'):
book_verses = nrsv_ot[nrsv_ot['_book'] == book][['osis_ref', '_num']].sort_values('_num')
if book_verses.empty:
continue
perps_sorted = book_perps[['_start_num', '_end_num', 'bt_layer', 'plate_segment']].sort_values('_start_num')
# merge_asof: for each verse, find the last pericope whose start ≤ verse
merged = pd.merge_asof(
book_verses,
perps_sorted,
left_on='_num',
right_on='_start_num',
direction='backward',
)
# Keep only verses that also fall before the pericope's end
merged = merged[merged['_num'] <= merged['_end_num']]
for _, row in merged.iterrows():
verse_bt[row['osis_ref']] = (int(row['bt_layer']), str(row['plate_segment']))
# ── 2. Single-verse nodes: unit_id IS the verse OSIS ref ───────────────
single_mask = dag['unit_id'].str.match(r'^\w+\.\d+\.\d+$', na=False)
for _, row in dag[single_mask][['unit_id', 'bt_layer', 'plate_segment']].iterrows():
verse_bt[row['unit_id']] = (int(row['bt_layer']), str(row['plate_segment']))
return verse_bt
def _normalize_osis(ref: str) -> str:
"""Convert lxx-style '1Chr 1:1' to OSIS dot-format '1Chr.1.1'."""
if ' ' in ref:
book, rest = ref.split(' ', 1)
return f"{book}.{rest.replace(':', '.')}"
return ref
def build_concept_verse_edge_bt() -> None:
"""Extend concept_verse_edge_sp with bt_layer and plate_segment from pericope-genealogy.
Cardinal rule (CLAUDE.md): preserves ALL verse-level rows (match_grain='verse').
bt_layer is nullable (pd.NA) for NT verses not represented as intertextuality
anchor nodes in composition_dag. osisRef is the normalized OSIS dot-format ref.
T8 dependency: reads NuBerea/pericope-genealogy composition_dag (T7). Falls back
to local build dir until genealogy HF upload is complete.
"""
import numpy as np
cve = load_hf_or_local("NuBerea/concept-analysis", "concept_verse_edge_sp").to_pandas()
dag = load_hf_or_local("NuBerea/analytics", "composition_dag").to_pandas()
nrsv = load_hf_or_local("NuBerea/nrsv").to_pandas()
verse_bt = _build_verse_bt_lookup(dag, nrsv)
# Normalize all osis_refs to dot format for lookup
cve['osisRef'] = cve['osis_ref'].apply(_normalize_osis)
bt_layers = []
plate_segments = []
for ref in cve['osisRef']:
val = verse_bt.get(ref)
if val is not None:
bt_layers.append(val[0])
plate_segments.append(val[1])
else:
bt_layers.append(pd.NA)
plate_segments.append(pd.NA)
cve['bt_layer'] = pd.array(bt_layers, dtype=pd.Int64Dtype())
cve['plate_segment'] = plate_segments
cve['match_grain'] = 'verse'
dst = _parquet_out("concept_verse_edge_bt")
cve.to_parquet(dst, index=False)
matched = cve['bt_layer'].notna().sum()
print(
f" concept_verse_edge_bt: {len(cve):,} rows, "
f"{matched:,} ({100*matched/len(cve):.1f}%) with bt_layer — "
f"T8 (pericope-genealogy)"
)
def build_concept_trajectory_bt() -> None:
"""Build concept × bt_layer centroids from concept_verse_edge_bt.
Cardinal rule: pericope-level derived view (match_grain='pericope') alongside
concept_trajectory_sp (3-way stratum). Never replaces it.
Only rows with a non-null bt_layer contribute. Concepts with fewer than 2
matched verses in a bt_layer are excluded (no meaningful centroid).
Schema mirrors concept_trajectory_sp with bt_layer replacing stratum, plus
match_grain='pericope' and osisRef (book prefix of contributing verses).
"""
import numpy as np
# Load from local build output (concept_verse_edge_bt must be built first)
bt_path = _parquet_out("concept_verse_edge_bt")
if not bt_path.exists():
raise RuntimeError(
"concept_verse_edge_bt not found — run build_concept_verse_edge_bt first."
)
cve_bt = pd.read_parquet(bt_path)
# Work only on rows with bt_layer assigned
matched = cve_bt[cve_bt['bt_layer'].notna()].copy()
matched['bt_layer'] = matched['bt_layer'].astype(int)
def _numpy_centroid(embeddings) -> list:
arr = np.stack([np.asarray(e) for e in embeddings])
c = arr.mean(axis=0)
return c.tolist()
def _mean_cosine(embeddings, centroid) -> float:
c = np.asarray(centroid)
sims = []
for e in embeddings:
v = np.asarray(e)
denom = np.linalg.norm(v) * np.linalg.norm(c)
if denom > 0:
sims.append(float(np.dot(v, c) / denom))
return float(np.mean(sims)) if sims else 0.0
rows = []
for (concept_id, bt_layer), grp in matched.groupby(['concept_id', 'bt_layer']):
if len(grp) < 2:
continue
embeddings = grp['embedding'].tolist()
centroid = _numpy_centroid(embeddings)
rows.append({
'concept_id': concept_id,
'concept_label': grp['concept_label'].iloc[0],
'bt_layer': int(bt_layer),
'verse_count': len(grp),
'centroid': centroid,
'mean_cosine_to_centroid': _mean_cosine(embeddings, centroid),
'verse_refs': sorted(set(grp['osisRef'].tolist())),
'model_id': grp['model_id'].iloc[0],
'match_grain': 'pericope',
})
df = pd.DataFrame(rows).sort_values(['concept_id', 'bt_layer']).reset_index(drop=True)
dst = _parquet_out("concept_trajectory_bt")
df.to_parquet(dst, index=False)
print(
f" concept_trajectory_bt: {len(df):,} rows "
f"({df['concept_id'].nunique()} concepts × {df['bt_layer'].nunique()} bt_layers) — "
f"T8 derived view, match_grain=pericope"
)
def _build_reference_centroids() -> "tuple[list, list, np.ndarray]":
"""Compute per-concept reference centroids pooled from LXX+NT+VG in SPhilBerta space.
Returns:
(concept_ids, concept_labels, centroid_unit_matrix)
centroid_unit_matrix shape: (n_concepts, 768), L2-normalized rows.
"""
import numpy as np
print(" Loading concept_verse_edge_sp for reference centroids...")
cve_sp = load_hf_or_local("NuBerea/concept-analysis", "concept_verse_edge_sp").to_pandas()
concept_ids = []
concept_labels_list = []
centroid_list = []
for cid, grp in cve_sp.groupby("concept_id"):
mats = np.stack([np.asarray(e, dtype=np.float32) for e in grp["embedding"]])
centroid_list.append(mats.mean(axis=0))
concept_ids.append(cid)
concept_labels_list.append(grp["concept_label"].iloc[0])
centroid_matrix = np.stack(centroid_list) # (n_concepts, 768)
norms = np.linalg.norm(centroid_matrix, axis=1, keepdims=True)
centroid_unit = centroid_matrix / (norms + 1e-8)
print(f" Reference space: {len(concept_ids)} concept centroids (768-dim SPhilBerta)")
return concept_ids, concept_labels_list, centroid_unit
def _project_pseu_stratum(
stratum: str,
concept_ids: list,
concept_labels_list: list,
centroid_unit: "np.ndarray",
) -> "pd.DataFrame":
"""Project one PSEU stratum onto pre-computed concept centroids.
Each verse is assigned to its single nearest concept (top-1 cosine similarity).
assignment_method='nearest_centroid' distinguishes these rows from topology-based
pericope_containment rows in concept_verse_edge_sp.
"""
import numpy as np
config_name = "deut_verse_embeddings_sp" if stratum == "deut" else "pseu_verse_embeddings_sp"
print(f" Loading {config_name} from NuBerea/features...")
pseu_df = load_hf_or_local("NuBerea/features", config_name).to_pandas()
print(f" Loaded {len(pseu_df):,} {stratum} verses")
emb_matrix = np.stack([np.asarray(e, dtype=np.float32) for e in pseu_df["embedding"]])
emb_norms = np.linalg.norm(emb_matrix, axis=1, keepdims=True)
emb_unit = emb_matrix / (emb_norms + 1e-8)
cos_scores = emb_unit @ centroid_unit.T # (n_verses, n_concepts)
best_idx = cos_scores.argmax(axis=1)
best_score = cos_scores[np.arange(len(pseu_df)), best_idx]
rows = []
for i, (_, verse_row) in enumerate(pseu_df.iterrows()):
ci = int(best_idx[i])
rows.append({
"concept_id": concept_ids[ci],
"concept_label": concept_labels_list[ci],
"osis_ref": verse_row["osis_ref"],
"stratum": stratum,
"edge_weight": float(best_score[i]),
"embedding": verse_row["embedding"],
"model_id": verse_row["model_id"],
"assignment_method": "nearest_centroid",
})
return pd.DataFrame(rows)
def build_concept_verse_edge_pseu() -> None:
"""Build concept_verse_edge_deut and concept_verse_edge_pseu.
Both use nearest-centroid projection from the pooled LXX+NT+VG concept
centroids in SPhilBerta space. PSEU pericopes are not yet represented in
concept-graph topology, so pericope_containment is not available for these
strata. assignment_method='nearest_centroid' records this distinction.
Loads the reference centroid matrix once and reuses it for both strata.
"""
import numpy as np
concept_ids, concept_labels_list, centroid_unit = _build_reference_centroids()
for stratum in ("deut", "pseu"):
config = f"concept_verse_edge_{stratum}"
df = _project_pseu_stratum(stratum, concept_ids, concept_labels_list, centroid_unit)
dst = _parquet_out(config)
df.to_parquet(dst, index=False)
concept_count = df["concept_id"].nunique()
mean_score = df["edge_weight"].mean()
print(
f" {config}: {len(df):,} rows "
f"({concept_count} unique concepts, mean cosine={mean_score:.4f})"
)
def build_concept_trajectory_pseu() -> None:
"""Build concept_trajectory_deut and concept_trajectory_pseu.
Aggregates concept_verse_edge_deut/pseu to per-concept centroids.
build_concept_verse_edge_pseu must be run first.
"""
import numpy as np
for stratum in ("deut", "pseu"):
edge_path = _parquet_out(f"concept_verse_edge_{stratum}")
if not edge_path.exists():
raise RuntimeError(
f"concept_verse_edge_{stratum} not found — run build_concept_verse_edge_pseu first."
)
cve = pd.read_parquet(edge_path)
rows = []
for cid, grp in cve.groupby("concept_id"):
mats = np.stack([np.asarray(e, dtype=np.float32) for e in grp["embedding"]])
centroid = mats.mean(axis=0)
cnorm = np.linalg.norm(centroid)
centroid_unit = centroid / (cnorm + 1e-8)
norms = np.linalg.norm(mats, axis=1, keepdims=True)
units = mats / (norms + 1e-8)
mean_cos = float((units @ centroid_unit).mean())
rows.append({
"concept_id": cid,
"concept_label": grp["concept_label"].iloc[0],
"stratum": stratum,
"verse_count": len(grp),
"centroid": centroid.tolist(),
"mean_cosine_to_centroid": mean_cos,
"verse_refs": grp["osis_ref"].tolist(),
"model_id": grp["model_id"].iloc[0],
"assignment_method": "nearest_centroid",
})
df = pd.DataFrame(rows).sort_values("concept_id").reset_index(drop=True)
config = f"concept_trajectory_{stratum}"
dst = _parquet_out(config)
df.to_parquet(dst, index=False)
mean_cos_all = df["mean_cosine_to_centroid"].mean()
print(
f" {config}: {len(df):,} rows "
f"({df['concept_id'].nunique()} concepts, mean coherence={mean_cos_all:.4f})"
)
def build_concept_drift_pseu() -> None:
"""Compute PSEU-extended drift metrics across 5 new stratum transition pairs.
Pairs: lxx→deut, lxx→pseu, deut→nt, pseu→nt, deut→pseu.
Mirrors the Schlattmann & Vogl 2024 methodology used in build_concept_drift(),
extending to LXX Deuterocanon (stratum 6) and Second Temple Pseudepigrapha (stratum 7).
Inputs:
NuBerea/concept-trajectories concept_trajectory_sp — lxx/nt/vg centroids
NuBerea/concept-trajectories concept_verse_edge_sp — lxx/nt/vg embeddings
NuBerea/concept-analysis concept_trajectory_deut — deut centroids (PSEU)
NuBerea/concept-analysis concept_trajectory_pseu — pseu centroids (PSEU)
NuBerea/concept-analysis concept_verse_edge_deut — deut embeddings (PSEU)
NuBerea/concept-analysis concept_verse_edge_pseu — pseu embeddings (PSEU)
Output: 78 rows (one per concept); wide format with per-stratum crystallization
and per-pair cross-stratum drift metrics.
"""
import numpy as np
sys.path.insert(0, str(_THIS.parent))
from analyze_concept_drift import cosine_sim, cross_density
traj_sp = load_hf_or_local("NuBerea/concept-analysis", "concept_trajectory_sp").to_pandas()
traj_deut = load_hf_or_local("NuBerea/concept-analysis", "concept_trajectory_deut").to_pandas()
traj_pseu = load_hf_or_local("NuBerea/concept-analysis", "concept_trajectory_pseu").to_pandas()
traj_all = pd.concat([traj_sp, traj_deut, traj_pseu], ignore_index=True)
traj_idx = traj_all.set_index(["concept_id", "stratum"])
cve_sp = load_hf_or_local("NuBerea/concept-analysis", "concept_verse_edge_sp").to_pandas()
cve_deut = load_hf_or_local("NuBerea/concept-analysis", "concept_verse_edge_deut").to_pandas()
cve_pseu = load_hf_or_local("NuBerea/concept-analysis", "concept_verse_edge_pseu").to_pandas()
cve_all = pd.concat([cve_sp, cve_deut, cve_pseu], ignore_index=True)
emb_index: dict[tuple, np.ndarray] = {}
for (cid, stratum), grp in cve_all.groupby(["concept_id", "stratum"]):
emb_index[(cid, stratum)] = np.stack(
[np.array(e, dtype=np.float32) for e in grp["embedding"]]
)
all_concepts = traj_sp["concept_id"].unique()
concept_label = (
traj_sp.drop_duplicates("concept_id")
.set_index("concept_id")["concept_label"]
.to_dict()
)
pseu_pairs = [
("lxx", "deut"), ("lxx", "pseu"),
("deut", "nt"), ("pseu", "nt"),
("deut", "pseu"),
]
all_strata = ["lxx", "nt", "vg", "deut", "pseu"]
records = []
for cid in all_concepts:
label = concept_label.get(cid, "")
row: dict = {"concept_id": cid, "concept_label": label}
for s in all_strata:
key = (cid, s)
if key in traj_idx.index:
row[f"verse_count_{s}"] = int(traj_idx.loc[key, "verse_count"])
row[f"crystallization_{s}"] = float(traj_idx.loc[key, "mean_cosine_to_centroid"])
row[f"dispersion_{s}"] = 1.0 - row[f"crystallization_{s}"]
else:
row[f"verse_count_{s}"] = 0
row[f"crystallization_{s}"] = float("nan")
row[f"dispersion_{s}"] = float("nan")
for s_a, s_b in pseu_pairs:
tag = f"{s_a}_{s_b}"
key_a = (cid, s_a)
key_b = (cid, s_b)
have_a = key_a in traj_idx.index
have_b = key_b in traj_idx.index
if have_a and have_b:
c_a = np.array(traj_idx.loc[key_a, "centroid"], dtype=np.float32)
c_b = np.array(traj_idx.loc[key_b, "centroid"], dtype=np.float32)
cc = cosine_sim(c_a, c_b)
row[f"centroid_cosine_{tag}"] = cc
row[f"centroid_drift_{tag}"] = 1.0 - cc
else:
row[f"centroid_cosine_{tag}"] = float("nan")
row[f"centroid_drift_{tag}"] = float("nan")
d_a = row[f"dispersion_{s_a}"]
d_b = row[f"dispersion_{s_b}"]
row[f"delta_dispersion_{tag}"] = (
d_b - d_a
if (not np.isnan(d_a) and not np.isnan(d_b))
else float("nan")
)
if have_a and key_b in emb_index:
c_a = np.array(traj_idx.loc[key_a, "centroid"], dtype=np.float32)
emb_b = emb_index[key_b]
cd_ab = cross_density(emb_b, c_a)
row[f"cross_density_{tag}"] = cd_ab
row[f"cross_drift_{tag}"] = 1.0 - cd_ab
else:
row[f"cross_density_{tag}"] = float("nan")
row[f"cross_drift_{tag}"] = float("nan")
if have_b and key_a in emb_index:
c_b = np.array(traj_idx.loc[key_b, "centroid"], dtype=np.float32)
emb_a = emb_index[key_a]
row[f"cross_density_{s_b}_{s_a}"] = cross_density(emb_a, c_b)
else:
row[f"cross_density_{s_b}_{s_a}"] = float("nan")
cd_tag = row[f"centroid_drift_{tag}"]
xd_tag = row[f"cross_drift_{tag}"]
both_ok = not np.isnan(cd_tag) and not np.isnan(xd_tag)
row[f"drift_score_{tag}"] = (cd_tag + xd_tag) / 2.0 if both_ok else float("nan")
records.append(row)
df = pd.DataFrame(records)
dst = _parquet_out("concept_drift_itp")
df.to_parquet(dst, index=False)
print(f" concept_drift_itp: {len(df)} rows, {len(df.columns)} cols — 5 PSEU drift pairs (lxx→deut/pseu, deut/pseu→nt, deut→pseu)")
def build_concept_aliases() -> None:
"""Build concept_aliases (TF-IDF English term → bt1_xxx mapping) from HF inputs.
Fully HF-native: loads bi_encoder_v1 + question_bin_assignment.json from
NuBerea/AI-models and L1 centroids from NuBerea/concept-graph, then runs the
BERTopic alias TF-IDF logic. See build_aliases_from_hf.py for details.
"""
from build_aliases_from_hf import build_aliases_dataframe
df = build_aliases_dataframe()
dst = _parquet_out("concept_aliases")
df.to_parquet(dst, index=False)
print(f" concept_aliases: {len(df)} rows, {len(df.columns)} cols — built HF-native")
def build_concept_verse_edge_patr() -> None:
"""Project patristic GRC segments onto concept centroids via nearest-centroid assignment.
Loads pre-computed SPhilBerta embeddings from NuBerea/features
(patr_grc_embeddings_sp, ~40K rows) and assigns each segment to its
nearest concept from the pooled LXX+NT+VG reference centroids.
Output schema matches concept_verse_edge_pseu but uses `segment_id` instead
of `osis_ref` and adds `cts_urn`, `author_label`, and `family_id` fields
for downstream patristic-specific joins.
"""
import numpy as np
concept_ids, concept_labels_list, centroid_unit = _build_reference_centroids()
print(" Loading patristics embeddings from NuBerea/features (patr_grc_embeddings_sp)...")
patr_df = load_hf_or_local("NuBerea/features", "patr_grc_embeddings_sp").to_pandas()
print(f" Loaded {len(patr_df):,} patristic GRC segments")
emb_matrix = np.stack([np.asarray(e, dtype=np.float32) for e in patr_df["embedding"]])
emb_norms = np.linalg.norm(emb_matrix, axis=1, keepdims=True)
emb_unit = emb_matrix / (emb_norms + 1e-8)
cos_scores = emb_unit @ centroid_unit.T # (n_segments, n_concepts)
best_idx = cos_scores.argmax(axis=1)
best_score = cos_scores[np.arange(len(patr_df)), best_idx]
rows = []
for i, (_, seg_row) in enumerate(patr_df.iterrows()):
ci = int(best_idx[i])
rows.append({
"concept_id": concept_ids[ci],
"concept_label": concept_labels_list[ci],
"segment_id": seg_row["segment_id"],
"cts_urn": seg_row["cts_urn"],
"author_label": seg_row["author_label"],
"family_id": seg_row["family_id"],
"stratum": "patr",
"edge_weight": float(best_score[i]),
"embedding": seg_row["embedding"],
"model_id": seg_row["model_id"],
"assignment_method": "nearest_centroid",
})
df = pd.DataFrame(rows)
dst = _parquet_out("concept_verse_edge_patr")
df.to_parquet(dst, index=False)
concept_count = df["concept_id"].nunique()
mean_score = df["edge_weight"].mean()
print(
f" concept_verse_edge_patr: {len(df):,} rows "
f"({concept_count} unique concepts, mean cosine={mean_score:.4f}, "
f"{df['family_id'].nunique()} families)"
)
_TARGUM_OSIS_MAP: dict[str, str] = {
# Torah
"Genesis": "Gen", "Exodus": "Exod", "Leviticus": "Lev",
"Numbers": "Num", "Deuteronomy": "Deut",
# Former Prophets
"Joshua": "Josh", "Judges": "Judg", "Ruth": "Ruth",
"I Samuel": "1Sam", "II Samuel": "2Sam",
"I Kings": "1Kgs", "II Kings": "2Kgs",
# Latter Prophets
"Isaiah": "Isa", "Jeremiah": "Jer", "Ezekiel": "Ezek",
"Hosea": "Hos", "Joel": "Joel", "Amos": "Amos", "Obadiah": "Obad",
"Jonah": "Jonah", "Micah": "Mic", "Nahum": "Nah", "Habakkuk": "Hab",
"Zephaniah": "Zeph", "Haggai": "Hag", "Zechariah": "Zech", "Malachi": "Mal",
# Writings
"Psalms": "Ps", "Proverbs": "Prov", "Job": "Job",
"Song of Songs": "Song", "Lamentations": "Lam", "Ecclesiastes": "Eccl",
"Esther": "Esth", "Daniel": "Dan", "Ezra": "Ezra", "Nehemiah": "Neh",
"I Chronicles": "1Chr", "II Chronicles": "2Chr",
}
def _embed_texts_sphilberta(texts: list[str], batch_size: int = 32) -> tuple:
"""Embed texts with SPhilBerta. Returns (vecs: np.ndarray float32, model_id: str)."""
from sentence_transformers import SentenceTransformer
import torch
model_id = "bowphs/SPhilBerta"
if torch.backends.mps.is_available():
device = "mps"
elif torch.cuda.is_available():
device = "cuda"
else:
device = "cpu"
print(f" Loading {model_id} on {device}...")
model = SentenceTransformer(model_id, device=device)
print(f" Embedding {len(texts):,} texts (batch_size={batch_size})...")
vecs = model.encode(
texts,
batch_size=batch_size,
show_progress_bar=True,
normalize_embeddings=False,
convert_to_numpy=True,
)
return vecs.astype("float32"), model_id
def build_concept_verse_edge_philo() -> None:
"""Project Philo of Alexandria GRC segments onto concept centroids.
Loads philo_grc from NuBerea/secondary-sources (7,906 rows), embeds with
SPhilBerta, and assigns each segment to its nearest concept centroid.
Schema mirrors concept_verse_edge_patr.
"""
import numpy as np
concept_ids, concept_labels_list, centroid_unit = _build_reference_centroids()
print(" Loading philo_grc from NuBerea/secondary-sources...")
philo_df = load_hf_or_local("NuBerea/secondary-sources", "philo_grc").to_pandas()
print(f" Loaded {len(philo_df):,} Philo segments")
vecs, model_id = _embed_texts_sphilberta(philo_df["text"].tolist())
norms = np.linalg.norm(vecs, axis=1, keepdims=True)
vecs_unit = vecs / (norms + 1e-8)
cos_scores = vecs_unit @ centroid_unit.T
best_idx = cos_scores.argmax(axis=1)
best_score = cos_scores[np.arange(len(philo_df)), best_idx]
rows = []
for i, (_, seg) in enumerate(philo_df.iterrows()):
ci = int(best_idx[i])
rows.append({
"concept_id": concept_ids[ci],
"concept_label": concept_labels_list[ci],
"segment_id": seg["segment_id"],
"cts_urn": seg["cts_urn"],
"author_label": seg["author_label"],
"work_label": seg["work_label"],
"stratum": "philo",
"edge_weight": float(best_score[i]),
"embedding": vecs[i],
"model_id": model_id,
"assignment_method": "nearest_centroid",
})
df = pd.DataFrame(rows)
dst = _parquet_out("concept_verse_edge_philo")
df.to_parquet(dst, index=False)
concept_count = df["concept_id"].nunique()
mean_score = df["edge_weight"].mean()
print(
f" concept_verse_edge_philo: {len(df):,} rows "
f"({concept_count} unique concepts, mean cosine={mean_score:.4f})"
)
def build_concept_verse_edge_josephus() -> None:
"""Project Josephus GRC segments onto concept centroids.
Loads josephus_grc from NuBerea/secondary-sources (12,735 rows), embeds with
SPhilBerta, and assigns each segment to its nearest concept centroid.
"""
import numpy as np
concept_ids, concept_labels_list, centroid_unit = _build_reference_centroids()
print(" Loading josephus_grc from NuBerea/secondary-sources...")
jos_df = load_hf_or_local("NuBerea/secondary-sources", "josephus_grc").to_pandas()
print(f" Loaded {len(jos_df):,} Josephus segments")
vecs, model_id = _embed_texts_sphilberta(jos_df["text"].tolist())
norms = np.linalg.norm(vecs, axis=1, keepdims=True)
vecs_unit = vecs / (norms + 1e-8)
cos_scores = vecs_unit @ centroid_unit.T
best_idx = cos_scores.argmax(axis=1)
best_score = cos_scores[np.arange(len(jos_df)), best_idx]
rows = []
for i, (_, seg) in enumerate(jos_df.iterrows()):
ci = int(best_idx[i])
rows.append({
"concept_id": concept_ids[ci],
"concept_label": concept_labels_list[ci],
"segment_id": seg["segment_id"],
"cts_urn": seg["cts_urn"],
"author_label": seg["author_label"],
"work_label": seg["work_label"],
"stratum": "josephus",
"edge_weight": float(best_score[i]),
"embedding": vecs[i],
"model_id": model_id,
"assignment_method": "nearest_centroid",
})
df = pd.DataFrame(rows)
dst = _parquet_out("concept_verse_edge_josephus")
df.to_parquet(dst, index=False)
concept_count = df["concept_id"].nunique()
mean_score = df["edge_weight"].mean()
print(
f" concept_verse_edge_josephus: {len(df):,} rows "
f"({concept_count} unique concepts, mean cosine={mean_score:.4f})"
)
def build_concept_verse_edge_targum() -> None:
"""Project Targum (Aramaic OT paraphrase) English translations onto concept centroids.
Loads targum config from NuBerea/second-temple (~28K rows), embeds text_en
with SPhilBerta (Aramaic is unvalidated; English preserves interpretive content),
and assigns each verse to its nearest concept centroid. Derives canonical osis_ref
from book + section_0 (chapter) + section_1 (verse) for downstream pericope joins.
"""
import numpy as np
concept_ids, concept_labels_list, centroid_unit = _build_reference_centroids()
print(" Loading targum from NuBerea/second-temple...")
tg_df = load_hf_or_local("NuBerea/second-temple", "targum").to_pandas()
print(f" Loaded {len(tg_df):,} Targum rows")
# Drop rows with no English translation
tg_df = tg_df[tg_df["text_en"].notna() & (tg_df["text_en"].str.strip() != "")].copy()
print(f" {len(tg_df):,} rows with text_en available")
texts = tg_df["text_en"].tolist()
vecs, model_id = _embed_texts_sphilberta(texts)
norms = np.linalg.norm(vecs, axis=1, keepdims=True)
vecs_unit = vecs / (norms + 1e-8)
cos_scores = vecs_unit @ centroid_unit.T
best_idx = cos_scores.argmax(axis=1)
best_score = cos_scores[np.arange(len(tg_df)), best_idx]
def _osis_book_from_targum(book: str, targum_name: str | None) -> str | None:
osis = _TARGUM_OSIS_MAP.get(book)
if osis:
return osis
# book is typically "{targum_name} {canonical_name}" — strip the prefix
prefix = (targum_name or "").strip()
if prefix and book.startswith(prefix + " "):
osis = _TARGUM_OSIS_MAP.get(book[len(prefix) + 1:])
if osis:
return osis
# Fallback: try each word-count suffix (handles unexpected prefix variants)
words = book.split()
for skip in range(1, len(words)):
osis = _TARGUM_OSIS_MAP.get(" ".join(words[skip:]))
if osis:
return osis
return None
rows = []
for i, (_, seg) in enumerate(tg_df.iterrows()):
ci = int(best_idx[i])
book = seg.get("book") or ""
targum_name = seg.get("targum_name")
osis_book = _osis_book_from_targum(book, targum_name)
s0 = seg.get("section_0")
s1 = seg.get("section_1")
if osis_book and s0 is not None and s1 is not None:
osis_ref = f"{osis_book}.{int(s0) + 1}.{int(s1) + 1}"
elif osis_book and s0 is not None:
osis_ref = f"{osis_book}.{int(s0) + 1}"
else:
osis_ref = None
rows.append({
"concept_id": concept_ids[ci],
"concept_label": concept_labels_list[ci],
"segment_id": seg["segment_id"],
"osis_ref": osis_ref,
"targum_name": seg.get("targum_name"),
"section": seg.get("section"),
"book": book,
"stratum": "targum",
"edge_weight": float(best_score[i]),
"embedding": vecs[i],
"model_id": model_id,
"assignment_method": "nearest_centroid",
})
df = pd.DataFrame(rows)
dst = _parquet_out("concept_verse_edge_targum")
df.to_parquet(dst, index=False)
concept_count = df["concept_id"].nunique()
mean_score = df["edge_weight"].mean()
osis_coverage = df["osis_ref"].notna().mean()
print(
f" concept_verse_edge_targum: {len(df):,} rows "
f"({concept_count} unique concepts, mean cosine={mean_score:.4f}, "
f"osis_ref coverage={osis_coverage:.1%})"
)
def build_concept_verse_edge_talmud() -> None:
"""Project Talmud English translations onto concept centroids.
Reads mishnah and tosefta directly from NuBerea/talmud via HfFileSystem
(load_dataset fails on this repo; parquets are read directly). Bavli and
yerushalmi are skipped until their parquets are uploaded. Embeds text_en
with SPhilBerta (Mishnaic Hebrew and Talmudic Aramaic are unembeddable
directly; English translations preserve interpretive content).
Stratum values: mishnah (~200 CE), tosefta (~220 CE).
"""
import numpy as np
from huggingface_hub import HfFileSystem
concept_ids, concept_labels_list, centroid_unit = _build_reference_centroids()
fs = HfFileSystem()
_TALMUD_PATHS = [
("datasets/NuBerea/talmud/data/mishnah/train-00000-of-00001.parquet", "mishnah"),
("datasets/NuBerea/talmud/data/tosefta/train-00000-of-00001.parquet", "tosefta"),
]
all_dfs: list[pd.DataFrame] = []
for hf_path, stratum in _TALMUD_PATHS:
try:
df = pd.read_parquet(fs.open(hf_path))
except Exception as e: # noqa: BLE001 best-effort catch (logs/records)
print(f" Skipping talmud/{stratum}: {e}")
continue
if len(df) == 0:
print(f" Skipping talmud/{stratum}: empty")
continue
total = len(df)
df = df[df["text_en"].notna() & (df["text_en"].str.strip() != "")].copy()
df["_stratum"] = stratum
print(f" talmud/{stratum}: {len(df):,} rows with text_en (of {total:,} total)")
all_dfs.append(df)
if not all_dfs:
print(" No Talmud data available — skipping")
return
combined = pd.concat(all_dfs, ignore_index=True)
print(f" Total: {len(combined):,} Talmud rows across {len(all_dfs)} config(s)")
texts = combined["text_en"].tolist()
vecs, model_id = _embed_texts_sphilberta(texts, batch_size=128)
norms = np.linalg.norm(vecs, axis=1, keepdims=True)
vecs_unit = vecs / (norms + 1e-8)
cos_scores = vecs_unit @ centroid_unit.T
best_idx = cos_scores.argmax(axis=1)
best_score = cos_scores[np.arange(len(combined)), best_idx]
rows = []
for i, (_, seg) in enumerate(combined.iterrows()):
ci = int(best_idx[i])
rows.append({
"concept_id": concept_ids[ci],
"concept_label": concept_labels_list[ci],
"segment_id": seg["segment_id"],
"ref": seg.get("ref"),
"title": seg.get("title"),
"tractate": seg.get("tractate"),
"order": seg.get("order"),
"stratum": seg["_stratum"],
"edge_weight": float(best_score[i]),
"embedding": vecs[i],
"model_id": model_id,
"assignment_method": "nearest_centroid",
})
df_out = pd.DataFrame(rows)
dst = _parquet_out("concept_verse_edge_talmud")
df_out.to_parquet(dst, index=False)
concept_count = df_out["concept_id"].nunique()
mean_score = df_out["edge_weight"].mean()
strata = df_out["stratum"].value_counts().to_dict()
print(
f" concept_verse_edge_talmud: {len(df_out):,} rows "
f"({concept_count} unique concepts, mean cosine={mean_score:.4f})"
)
print(f" Strata: {strata}")
def build_concept_verse_edge_nhc() -> None:
"""Project Nag Hammadi Corpus segments onto concept centroids.
Reuses pre-computed SPhilBerta embeddings from NuBerea/features →
gnostic_ln_bridge rather than re-running inference. The embedding
column is a 768-dim float32 vector already L2-normalised at build time.
"""
import numpy as np
from datasets import load_dataset
concept_ids, concept_labels_list, centroid_unit = _build_reference_centroids()
ds = load_dataset("NuBerea/features", "gnostic_ln_bridge", split="train")
df = ds.to_pandas()
emb_col = df["embedding"].tolist()
vecs = np.array(emb_col, dtype=np.float32)
norms = np.linalg.norm(vecs, axis=1, keepdims=True)
vecs_unit = vecs / (norms + 1e-8)
cos_scores = vecs_unit @ centroid_unit.T
best_idx = cos_scores.argmax(axis=1)
best_score = cos_scores[np.arange(len(df)), best_idx]
rows = []
for i, (_, seg) in enumerate(df.iterrows()):
ci = int(best_idx[i])
rows.append({
"concept_id": concept_ids[ci],
"concept_label": concept_labels_list[ci],
"segment_id": seg["segment_id"],
"codex": seg.get("codex"),
"tractate_num": seg.get("tractate_num"),
"work_slug": seg.get("work_slug"),
"work_title": seg.get("work_title"),
"stratum": "nhc",
"edge_weight": float(best_score[i]),
"embedding": vecs[i],
"model_id": "bowphs/SPhilBerta",
"assignment_method": "nearest_centroid",
})
df_out = pd.DataFrame(rows)
dst = _parquet_out("concept_verse_edge_nhc")
df_out.to_parquet(dst, index=False)
concept_count = df_out["concept_id"].nunique()
mean_score = df_out["edge_weight"].mean()
print(
f" concept_verse_edge_nhc: {len(df_out):,} rows "
f"({concept_count} unique concepts, mean cosine={mean_score:.4f})"
)
# ──────────────────────────────────────────────────────────────────────────────
# concept_trajectory_series + concept_drift_sequential
# Corpus-timeline view: per-concept centroids across 11 strata in temporal order
# (OT-LXX → PSEU → Philo/Josephus → NT → Patristics → Targum/Mishnah/Tosefta → VG)
# ──────────────────────────────────────────────────────────────────────────────
_SERIES_STRATA = [
# (stratum_name, stratum_order, language, edge_config, stratum_filter)
# stratum_filter: column value to select when one config covers multiple strata
("lxx", 1, "grc", "concept_verse_edge_sp", "lxx"),
("deut", 2, "grc", "concept_verse_edge_deut", None),
("pseu", 3, "grc", "concept_verse_edge_pseu", None),
("philo", 4, "grc", "concept_verse_edge_philo", None),
("nt", 5, "grc", "concept_verse_edge_sp", "nt"),
("josephus", 6, "grc", "concept_verse_edge_josephus", None),
("patr", 7, "grc", "concept_verse_edge_patr", None),
("targum", 8, "en", "concept_verse_edge_targum", None),
("mishnah", 9, "en", "concept_verse_edge_talmud", "mishnah"),
("tosefta", 10, "en", "concept_verse_edge_talmud", "tosefta"),
("vg", 11, "lat", "concept_verse_edge_sp", "vg"),
]
_SERIES_PAIRS = [
# (stratum_a, stratum_b, pair_order)
# patr→targum (pair 7) and tosefta→vg (pair 10) are cross-language
("lxx", "deut", 1),
("deut", "pseu", 2),
("pseu", "philo", 3),
("philo", "nt", 4),
("nt", "josephus", 5),
("josephus", "patr", 6),
("patr", "targum", 7),
("targum", "mishnah", 8),
("mishnah", "tosefta", 9),
("tosefta", "vg", 10),
]
def build_concept_trajectory_series() -> None:
"""Per-concept centroids across all 11 strata in temporal order.
Long format: ~858 rows (78 concepts × up to 11 strata).
Crystallization = mean cosine of member embeddings to their own per-stratum centroid.
mean_edge_weight = mean cosine to the reference centroid (pooled LXX+NT+VG).
Strata with no segments for a concept are omitted (so row count may be <858).
concept_verse_edge_sp is loaded once and reused for lxx, nt, vg strata.
concept_verse_edge_talmud is loaded once and reused for mishnah, tosefta strata.
"""
import numpy as np
_cache: dict[str, pd.DataFrame] = {}
all_rows: list[dict] = []
for stratum, stratum_order, language, edge_config, stratum_filter in _SERIES_STRATA:
print(f" [{stratum_order}/11] {stratum} ← {edge_config}...")
if edge_config not in _cache:
_cache[edge_config] = load_hf_or_local(
"NuBerea/concept-analysis", edge_config
).to_pandas()
df = _cache[edge_config].copy()
if stratum_filter is not None:
df = df[df["stratum"] == stratum_filter]
if len(df) == 0:
print(f" Skipping {stratum}: no rows")
continue
model_id = str(df["model_id"].iloc[0]) if "model_id" in df.columns else "SPhilBerta"
for cid, grp in df.groupby("concept_id"):
embs = np.stack([np.asarray(e, dtype=np.float32) for e in grp["embedding"]])
centroid = embs.mean(axis=0)
cnorm = np.linalg.norm(centroid)
centroid_unit = centroid / (cnorm + 1e-8)
norms = np.linalg.norm(embs, axis=1, keepdims=True)
units = embs / (norms + 1e-8)
crystallization = float((units @ centroid_unit).mean())
all_rows.append({
"concept_id": cid,
"concept_label": grp["concept_label"].iloc[0],
"stratum": stratum,
"stratum_order": stratum_order,
"language": language,
"verse_count": len(grp),
"mean_edge_weight": float(grp["edge_weight"].mean()),
"crystallization": crystallization,
"centroid": centroid.tolist(),
"model_id": model_id,
})
n_concepts = df["concept_id"].nunique()
print(f" {stratum}: {len(df):,} segments → {n_concepts} concept groups")
result = (
pd.DataFrame(all_rows)
.sort_values(["concept_id", "stratum_order"])
.reset_index(drop=True)
)
dst = _parquet_out("concept_trajectory_series")
result.to_parquet(dst, index=False)
print(
f" concept_trajectory_series: {len(result):,} rows "
f"({result['concept_id'].nunique()} concepts × "
f"{result['stratum'].nunique()} strata)"
)
def build_concept_drift_sequential() -> None:
"""Adjacent-pair drift across the 11-stratum corpus timeline.
Reads concept_trajectory_series (must be built first).
Outputs ~780 rows (78 concepts × 10 adjacent pairs).
same_language=False marks cross-language pairs (patr→targum, tosefta→vg)
where centroid_drift is confounded by model-language mismatch.
centroid_drift = 1 − cosine(centroid_a, centroid_b); range [0, 2].
delta_crystallization = crystallization_b − crystallization_a.
delta_mean_edge_weight = mean_edge_weight_b − mean_edge_weight_a.
"""
import numpy as np
series_path = _parquet_out("concept_trajectory_series")
if not series_path.exists():
raise RuntimeError(
"concept_trajectory_series not found — "
"run build_concept_trajectory_series first."
)
traj = pd.read_parquet(series_path)
traj_idx = traj.set_index(["concept_id", "stratum"])
lang_map = traj.drop_duplicates("stratum").set_index("stratum")["language"].to_dict()
concept_ids = traj["concept_id"].unique()
all_rows: list[dict] = []
for stratum_a, stratum_b, pair_order in _SERIES_PAIRS:
lang_a = lang_map.get(stratum_a, "?")
lang_b = lang_map.get(stratum_b, "?")
for cid in concept_ids:
try:
row_a = traj_idx.loc[(cid, stratum_a)]
row_b = traj_idx.loc[(cid, stratum_b)]
except KeyError:
continue
c_a = np.asarray(row_a["centroid"], dtype=np.float32)
c_b = np.asarray(row_b["centroid"], dtype=np.float32)
c_a_u = c_a / (np.linalg.norm(c_a) + 1e-8)
c_b_u = c_b / (np.linalg.norm(c_b) + 1e-8)
cos_sim = float(np.dot(c_a_u, c_b_u))
all_rows.append({
"concept_id": cid,
"concept_label": row_a["concept_label"],
"stratum_a": stratum_a,
"stratum_b": stratum_b,
"pair_order": pair_order,
"language_a": lang_a,
"language_b": lang_b,
"same_language": lang_a == lang_b,
"centroid_cosine": cos_sim,
"centroid_drift": 1.0 - cos_sim,
"delta_crystallization": float(row_b["crystallization"] - row_a["crystallization"]),
"delta_mean_edge_weight": float(row_b["mean_edge_weight"] - row_a["mean_edge_weight"]),
})
result = (
pd.DataFrame(all_rows)
.sort_values(["concept_id", "pair_order"])
.reset_index(drop=True)
)
dst = _parquet_out("concept_drift_sequential")
result.to_parquet(dst, index=False)
n_pairs = result["pair_order"].nunique()
same_lang_count = int(result["same_language"].sum())
print(
f" concept_drift_sequential: {len(result):,} rows "
f"({n_pairs} pairs × ~{len(result) // max(n_pairs, 1)} concepts; "
f"{same_lang_count} same-language rows)"
)
CONFIGS = {
"concept_verse_edge": lambda: import_trajectory_config("concept_verse_edge"),
"concept_trajectory": lambda: import_trajectory_config("concept_trajectory"),
"concept_verse_edge_sp": lambda: import_trajectory_config("concept_verse_edge_sp"),
"concept_trajectory_sp": lambda: import_trajectory_config("concept_trajectory_sp"),
"concept_verse_edge_bt": build_concept_verse_edge_bt,
"concept_trajectory_bt": build_concept_trajectory_bt,
"concept_verse_edge_pseu": build_concept_verse_edge_pseu,
"concept_trajectory_pseu": build_concept_trajectory_pseu,
"concept_drift": build_concept_drift,
"concept_drift_itp": build_concept_drift_pseu,
"concept_drift_hamilton": build_concept_drift_hamilton,
"community_candidate": lambda: import_community_config("community_candidate"),
"community_membership": lambda: import_community_config("community_membership"),
"concept_centroids": build_concept_centroids,
"concept_aliases": build_concept_aliases,
"concept_verse_edge_patr": build_concept_verse_edge_patr,
"concept_verse_edge_philo": build_concept_verse_edge_philo,
"concept_verse_edge_josephus": build_concept_verse_edge_josephus,
"concept_verse_edge_targum": build_concept_verse_edge_targum,
"concept_verse_edge_talmud": build_concept_verse_edge_talmud,
"concept_verse_edge_nhc": build_concept_verse_edge_nhc,
"concept_trajectory_series": build_concept_trajectory_series,
"concept_drift_sequential": build_concept_drift_sequential,
}
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--config", help="Build only this config")
args = parser.parse_args()
configs = {args.config: CONFIGS[args.config]} if args.config else CONFIGS
if args.config and args.config not in CONFIGS:
print(f"Unknown config '{args.config}'. Valid: {list(CONFIGS)}")
raise SystemExit(1)
print(f"Building {len(configs)} concept-analysis config(s) ...")
for name, fn in configs.items():
fn()
print(f"\nDone. Output in {OUT}")
if __name__ == "__main__":
main()
|