isnad-chrono v0.1.1

Summary

A corpus-wide isnad-cum-matn dataset derived from hdith.com: every parsed transmission chain, the takhrij groups that bundle the versions of one report, merged chain trees with their common links (k = 2 and k = 3), inferred death intervals for every narrator, report dates, and the attribution of wording changes to the transmitter who introduced them. The release contains derived numbers, ids, short categorical labels and 16-hex text hashes only. No hadith text, matn, narrator name or critic statement is redistributed.

• Source: hdith.com (narrator ids are hdith.com p-N slugs, hadith ids are hdith.com entry ids, group ids are hdith.com takhrij group slugs). Years are Hijri.
• License of the derived numbers: CC-BY-4.0 with attribution to hdith.com as the data source. The license of hdith.com itself is unstated; the text stays on the site.
• Version: v0.1.1. complete on a group means every occurrence page was cached when the groups stage ran; all_occurrences_have_chain means every occurrence also has a parsed chain. A later rebuild after the full crawl will update both.

Changelog

• v0.1 → v0.1.1: dating cleans repeated-narrator chains (as the trees do), edit-rate denominator restricted to the groups that have rated edits, complete-flag wording plus the new all_occurrences_have_chain flag, common-link date method (cl_k{k}_date_method).

Release v0.1.1: partial crawl

This release was built from a partial crawl of hdith.com: 103022 hadith pages were cached when the release was built (the chain and group stages ran on a slightly earlier snapshot of the same crawl while it was still growing). The Six Books are complete (0.9997 of their takhrij-group occurrences cached). The other collections are partial or not yet crawled. As a result:

• complete groups (every occurrence page cached when the groups stage ran): 0.465 of 30959 groups;
• groups whose every occurrence has a parsed chain (all_occurrences_have_chain): 0.432;
• occurrences with a parsed chain: 0.413;
• edit rows with status no_chain (no carrier has a parsed chain, mostly carriers in collections not yet crawled): 0.519.

Trees, common links and edit attributions of incomplete groups can change. v0.2 will be rebuilt from the full crawl.

Configs

config	rows	content
narrators	23487	one row per narrator: recorded dates and inferred intervals
chains	702156	one row per narrator position per chain
groups	30959	one row per takhrij group
occurrences	209124	group to hadith membership
tree_nodes	340067	one row per narrator per group component
report_dates	33587	one row per (group, root Companion)
edits	157560	one row per wording change (substitution segment)
edge_scores	190254	v1 per-edge chronology scores (Six Books)
checks	1	one row: the automatic checks below

Optional configs, present when their stage output exists: segments (alignment segments as hashes), shawahid (corroborating reports by other Companions, ids only), shawahid_companions, edit_rates (per-generation edit rates), dropped_links (links removed by the dating step).

Columns

• narrators: rawi_id; generation_num (the site's tabaqa number); death_year, death_kind, birth_year, birth_kind (recorded dates and their kind: exact, alt, approx, after, before, era, none); death_lo, death_hi, birth_lo, birth_hi (inferred interval, Hijri); support (number of distinct surviving teacher-student links touching the narrator); method (known, inferred, prior_only).
• chains: hadith_id, chain_no (one chain per path to the compiler after splitting at ح), position (0 = earliest narrator, the compiler is last), rawi_id, term (the transmission formula by which the next narrator received from this one, e.g. حدثنا, عن; a term longer than 6 words is a scribal note rather than a formula and is withheld as null: 5 cells across chains and edge_scores), unjoined (an earlier tahwil segment that could not be joined to the final chain), source (html, flat, html+flat).
• groups: group_id, n_entries, n_chains (the site's counts), books, n_occurrences, complete (every occurrence page was cached when the groups stage ran), all_occurrences_have_chain (every occurrence hadith has a parsed, joined chain in chains; computed by the tree stage).
• occurrences: group_id, hadith_id, book_slug, numbering (the hadith number in its book).
• tree_nodes: group_id, root_id (the Companion of the component), rawi_id, depth (BFS distance from the root), k_branches (children whose subtree reaches a compiler), n_descendant_chains, is_cl_k2, is_cl_k3 (common link), is_pcl_k2, is_pcl_k3 (partial common link).
• report_dates: group_id, root_id, for k in {2, 3}: cl_k{k}_id, cl_k{k}_date_lo, cl_k{k}_date_hi (death interval of the common link = the report date), cl_k{k}_date_method (the common link's dating method: known, inferred, or prior_only = no chain constraint reached it, so the date is the prior alone), cl_k{k}_all_chains (every chain passes through it); n_chains, n_compilers, complete, all_occurrences_have_chain (copied from the group).
• edits: group_id, seg_no, introduced_by_rawi_id, generation_depth, edit_type (sub, ins, del), status (attributed, ambiguous, partial, cross_root, no_chain, no_common_node), ambiguous, is_leaf_edit (single-witness variant at the compiler), has_tahwil_carrier, n_carriers, n_carriers_with_chain, n_base_carriers.
• edit_rates: depth, n_transmissions, n_edits, rate, n_edits_nonleaf, rate_nonleaf, from attributed and ambiguous rows only. The denominator is restricted the same way as the numerator: n_transmissions counts the chain links of only those groups that have at least one attributed or ambiguous row. Depth 0 is the Companion: edits attributed there have no transmission into the node, so n_transmissions is 0 and the rates are null.
• shawahid: group_id, hadith_ids (the corroborating reports). In this release companion_rawi_id was null on every row (the site sends no narrator for shawahid groups), so the column is dropped from shawahid. For the Companion of each corroborating report use shawahid_companions (group_id, hadith_id, companion_rawi_id), resolved from the parsed chains: 0.384 of its rows have a Companion.
• edge_scores: as in v1 (teacher_id, student_id, hadith_id, position, term, p_meet, evidence flags, verdict).
• checks: flattened keys, prefixed alignment_, sanity_, coverage_, dating_, plus the unprefixed release-status keys n_hadith_pages_cached, six_books_cached_rate, groups_complete_rate, all_occurrences_have_chain_rate, occurrence_chain_rate, edits_no_chain_rate, cl_k2_rate_multi, cl_k3_rate_multi, cl_k2_bounded_rate, cl_k2_prior_only_rate.

Method

1. Chains. Parsed from the annotated isnad HTML of each hadith page (narrator spans and formula spans in reading order). The chain is split at every ح (tahwil), at و-formulas (وحدثنا, وأخبرنا, ...) and at "وقال NAME:" restarts that name an earlier narrator (the new segment replays the path down to that narrator); each earlier segment is joined to the final chain at its last shared narrator, giving one chain per path to the compiler. A join phrase (كلاهما, كلهم, جميعا, ثلاثتهم) before "عن X" makes every open segment continue through X. Two adjacent narrator spans are parallel (and fork the chain) only when joined by و; adjacent spans with no و are linked in sequence, with the connective in the gap (أن, عن, ...) as the term. Chains that stop short of the flat isnad's earliest narrators are completed from it.
2. Groups. Taken from the site's takhrij group of each page: occurrences, and the site's variant segments (base text plus substitutions with their carrier hadiths), kept as hashes.
3. Narrator dating. Variables birth and death per narrator; each link student <- teacher gives birth(student) + 8 <= death(teacher); lifespans are bounded to 20..100 years; recorded dates enter as tight intervals. Chains are cleaned as for the trees (a chain is cut before its first repeated narrator) before the links are read. Difference constraints are propagated over the narrator graph. When a narrator's interval empties in a round, one link is dropped for it (listed in dropped_links): among the links that bound it that round, a link whose removal alone restores feasibility is preferred, then an ambiguous-class link (عن, أن, ...) before an explicit one (حدثنا, ...), then the lowest weight (fewest chains). Accuracy is measured by a 5-fold hold-out of recorded death years.
4. Trees and common links. All chains of a group's occurrences are merged into one graph per root Companion. The common link under k is the shallowest node with at least k branches reaching a compiler whose subtree contains every compiler; other nodes with at least k branches are partial common links. The report date is the common link's inferred death interval.
5. Edit attribution. For each substitution, the narrator that lies on every chain of every carrier and is deepest in the tree introduced the change; ambiguous when that narrator also transmitted the base text. Rows with status partial (some carriers without a parsed chain), cross_root, no_chain or no_common_node are excluded from the edit rates, which count only attributed and ambiguous rows. An edit's tree is built from all the group's witnesses, including substitution carriers that are missing from the group's occurrence list (such a carrier is never counted as a base carrier).

Checks

Computed when the release is built and shipped in the checks config and checks.json.

• Alignment check: the site's variant segments were checked against each carrier's own text in 500 sampled groups (425 with carrier text available; 6347 scored (substitution, carrier) pairs, 10251 pairs without cached carrier text).

  • Surface: the substitution's wording occurs in the carrier's text in 0.973 of pairs, against 0.001 in texts of the same group that do not carry it (control).
  • Position: re-aligning the base reading against the carrier text with difflib on token lists, a local difference (at most the substitution length + 3 tokens) lies within ±1 token of the site's span in 0.468 of pairs, against 0.203 at the same slots for texts without the variant (25742 control pairs). Carrier texts come from other books and editions and differ from the base reading in many places, so the positional agreement is indicative only.

• Narrator dating hold-out (median absolute error in years / mean absolute error / 90% interval coverage / median width), n = 6352:

  • propagation: 11.500 / 36.237 / 0.904 / 68.000
  • propagation, bounded intervals only (width ≤ 300 years, n = 6151): 11.000 / 14.578 / 0.909
  • baseline by chain depth: 19.146 / 24.293 / 0.903
  • baseline by generation mean: 14.166 / 18.370 / 0.885

  The median error of propagation is lower than the depth baseline's (11.5 vs 19.1 years) and its mean error is higher than the baseline's (36.2 vs 24.3). The mean is pulled up by a tail of large errors from narrators whose inferred interval is unbounded (no constraint closes it); restricted to bounded intervals the mean error is 14.6 years.

• Dropped links (negative cycles): 2584.

• Tree sanity: compilers that are leaves 0.973 (1571 of 58262 compilers also narrate inside another chain of the same component); roots without a parent in the group graph 0.974; root death interval starting no later than every dated compiler's ends (spec "root earliest") 0.997 of 32145 dated components.

• Coverage: 108240 chains (1799 unjoined); occurrences with a parsed chain 0.413; complete groups 0.465; edit rows in the rates (status attributed or ambiguous) 0.256.

• Common links: among the 12699 groups with at least 2 parsed chains, 0.896 have a k=2 common link and 0.375 a k=3 one; among all 30959 groups (including those with 0 or 1 chain) the rates are 0.368 and 0.154. Every k=2 common link has an inferred death interval; 0.992 of the 12066 k=2 common links have a bounded one (width ≤ 300 years). 20 k=2 common links (0.002) are dated by the prior alone (cl_k2_date_method = prior_only: no chain constraint reached the narrator); for k=3, 7 of 4924 (0.001). Filter on cl_k{k}_date_method before using report dates.

Limitations

• The site's alignment is taken as given and only spot-checked (see Checks); the site places a substitution after the base segment it replaces the end of, so the exact replaced span is not known and long reports align partially.
• ح-splitting depends on the HTML markers; unjoined segments are flagged, not guessed.
• Narrators far from any recorded date get wide intervals; method = prior_only means no chain constraint reached them.
• A common link is a structural property of the parsed chains; it does not by itself establish who put a report into circulation.
• Groups with complete = false lack some occurrences; their trees, common links and edit attributions can change after the full crawl.
• Edit attribution uses every parsed chain of a carrier; tahwil carriers are flagged (has_tahwil_carrier) for sensitivity analysis. Edit rates are edits per transmission, not probabilities.
• Chronology and common-link results are not authenticity verdicts.
• The spec's earliest_text_hash of report_dates is deferred to v0.2 (not computed in v0.1 or v0.1.1).
• Text hashes hide the text but are not a secrecy mechanism: a short segment (one or two words) can be confirmed by hashing candidate wordings.

Source and license

Derived from hdith.com. No source text is redistributed: segment texts appear only as 16-hex SHA-256 prefixes of the normalised text. The derived numbers are released under CC-BY-4.0; attribution to hdith.com as the data source is required. The license of hdith.com is unstated.

Citation

@misc{isnadchrono2026,
  title  = {isnad-chrono: a corpus-wide isnad-cum-matn dataset},
  author = {Haroon, Hesham},
  year   = {2026},
  note   = {Version v0.1.1. Derived from hdith.com},
  url    = {https://huggingface.co/datasets/HeshamHaroon/isnad-chrono}
}