buddhomics treats the Pali canon as one instrument: a place to browse by topic, follow a concept to the texts that carry it, read those texts in Pali and English side by side, and look up any word — all against a map the machine drew from the vocabulary alone.
The corpus
Everything is built on bilara-data, SuttaCentral's segment-aligned edition of the canon: the Pali root text of the Mahāsaṅgīti Tipiṭaka (the Sixth-Council recension), split into ~447,000 numbered segments across 6,959 documents, each Pali segment paired with its English translations where they exist. Because the text is segment-aligned, a translation can be laid exactly over the Pali — which is what makes a shared feature space, and the reader's line-by-line view, possible. Segment-aligned English exists for a little under half the canon, contributed by SuttaCentral translators — chiefly Bhikkhu Sujato and Bhikkhu Brahmali, with Bhikkhu Bodhi's numbering, and partial renderings by John Kelly, Bhikkhu Soma, Bhikkhu Suddhāso, and others. Where more than one translator has rendered a text, all of them are kept and selectable in the reader.
Machine translation of the untranslated remainder
The half of the canon without segment-aligned English is not obscure filler — it is the Abhidhamma (the entire Paṭṭhāna, Yamaka, Kathāvatthu…), the dense late Khuddaka books (Niddesa, Paṭisambhidāmagga, the Apadāna and Buddhavaṁsa), and stretches of Vinaya. To make the whole map legible and searchable in English, the untranslated remainder — about 216,000 segments across 2,608 documents — was machine-translated with Claude Fable 5. We chose it by translating a set of the canon's hardest passages with several leading models — Claude Opus 4.8 and Sonnet 5, and OpenAI's GPT-5.6 — and comparing the results side by side (read the comparison). Using one model throughout keeps the English style consistent across the corpus. A handful of passages that Fable would not complete were finished with GPT-5.6 instead.
These are unedited machine renderings, offered as a research aid, never as
published translations — the reader labels them plainly and marks each machine segment's
model. They are also released in bilara/sc-data format for anyone who finds them useful:
chercherai/bilara-data
(branch machine-translations).
How the maps are made
Each of the 6,959 texts is turned into a vector of TF-IDF weights over a ~40,000-term Pali vocabulary (tokens, tag-stripped), and that texts-by-terms matrix is reduced to 100 dimensions with truncated SVD — a single latent-semantic space that gives coordinates to both texts and terms. The texts map lays out the document vectors and the concept map lays out the term vectors, each with t-SNE; points that sit near each other share vocabulary. The colours are the traditional basket/collection groupings — the layout knows nothing about them, which is exactly what makes the agreements and the surprising mixtures worth looking at. Cluster labels come from single-linkage clustering on the t-SNE points, cut at several heights and named by a pass of language-model labelling.
How the panels link
A term–document incidence matrix over the full Pali text joins the two maps. Click a concept to light every text that contains it; click a text to see its vocabulary footprint on the concept map and open it in the reader; click any Pali word in the reader to look it up and find it among the concepts. A curated taxonomy down the left — 339 entries across People, Places, Beings, Cosmology, Rhetoric, Numbered Lists, Doctrine, Abhidhamma, Vinaya, Narrative, and Translations — filters both maps at once; each entry is a set of Pali stem-patterns validated against the corpus.
Reader, dictionary, and search
The reader shows Pali and English by segment, with a Pali / Both / English toggle and,
when a text has more than one translation, an independent toggle per translator (human =
circle, machine = square) so you can read any one or several at once. Word lookup layers
three dictionaries: the New Concise Pali English
Dictionary for concise prose, the Digital Pāḷi
Dictionary (inflected-form keyed, ~142k entries) for grammar and homonyms, and the
Dictionary of Pāḷi Proper Names for people and places. Search is accent-insensitive
(metta finds mettā) and runs over references, English full-text, and
a lemma-collapsed Pali full-text index, so a search folds every inflected form of a word
together.
Printing and PDFs
Every text has a PDF link that opens it as a typeset page — segments stitched back into flowing paragraphs, verse kept as verse, whichever translations you had selected — which your browser can print or save as a PDF. Where a group is small enough (all 34 Dīgha discourses, say) the whole group can be printed as one document with a contents page. The Pali and the human translations are CC0 from SuttaCentral, and our machine translations are offered as CC0 too, so the resulting files are yours to copy and give away. The commentaries are the exception: their text and translation are under unresolved licensing, so they can be read here but not exported.
Sources & code
- chercherai/buddhomics
- This project — analysis pipeline and the static site. Runs locally, deploys as flat files.
- chercherai/bilara-data
- Our machine translations in bilara/sc-data format (branch
machine-translations). - suttacentral/bilara-data
- The segment-aligned root text and human translations. Mahāsaṅgīti Pali; English by Sujato, Brahmali, et al.
- suttacentral/sc-data
- Dictionaries and reference data — NCPED, DPPN, and more — via SuttaCentral.
- Digital Pāḷi Dictionary (DPD)
- Bhikkhu Bodhirasa's inflected-form dictionary; the reader's grammar and homonym source.
- SuttaCentral
- The home of the underlying edition and translations; every text links back to its page there.
Caveats
Lexical similarity is not doctrine, chronology, or authorship. t-SNE distorts global distances by design — read neighbourhoods, not absolute positions. Cluster centroids blur internal diversity. Machine translations carry the errors of machine translation, especially in the technical Abhidhamma. Treat everything here as a lens for finding questions, not answers.