Opening on Vijayadaśamī, 20 October 2026, the traditional day to begin learning.
Read and learn Sanskrit, with AI doing the reading and Pāṇini's rules checking every word.
Read Sanskrit texts such as the Bhagavad Gītā and Kālidāsa's Meghadūta word by word, with an English translation of every verse, or paste a verse of your own. Every word's analysis is checked against Pāṇini's rules. Tap any word to see why it is what it is.
For anyone learning or teaching Sanskrit, and for people building and testing AI. New to Sanskrit? Start here.
The Digital Corpus of Sanskrit records each word's dictionary form and grammar by hand: adhikāraḥ → adhikāra, prathamā, puṃliṅga, ekavacana.
Vidyut, a working implementation of more than 2,000 sūtras of the Aṣṭādhyāyī, tests whether its rules allow that analysis for this form. Where they do, it also derives the word sūtra by sūtra. The rules say what a form can be; choosing among the possibilities in context is still the scholar's reading.
The rules allow the analysis. The rules analyse the form differently, so it is open for review. The rules have no entry yet. A compound member, not yet checked.
Compound members are not yet checked, so they are left out of the agreement figure.
An open test of how well AI models know Sanskrit grammar, with every answer checked by rule.
Every answer is an exact form or analysis, checked by rule. The answer key comes from Pāṇini's rules and a scholarly corpus, reviewed by hand. Items and grader are open. How the benchmark works
| Task | What the model must do | Items |
|---|---|---|
| Tiṅanta | Produce a verb form from root, lakāra, person and number | 150 |
| Subanta | Decline a stem in a given case and number | 94 |
| Analysis | Identify lakāra, puruṣa and vacana of a verb form | 100 |
| Gītā in context | Give a word's lemma and grammar inside a verse | 93 |
| Chandas | Scan a half-verse into guru and laghu syllables | 50 |
Whole verses from texts the models can't recite, plus rare verb forms. The test questions are private and the answer keys are under review. Method and the recitation test
| Model | Overall | Padapāṭha | Parse | Scan | Generate |
|---|
One word or form at a time, mostly from the Bhagavad Gītā. Method, sample questions and the full item set
| Model | Overall | Tiṅanta | Subanta | Analysis | Gītā | Chandas |
|---|
One email when there is something real to show. No sharing, and you can ask to be removed any time.
Word analyses: the Digital Corpus of Sanskrit by Oliver Hellwig, licensed CC BY 4.0. Its Gītā follows the critical edition of the Mahābhārata; verse numbers here use the common numbering, which differs in chapter 13. Meghadūta verses are numbered as in the corpus, counted separately in the Pūrvamegha and Uttaramegha.
Grammar engine: Vidyut from the Ambuda project, which implements more than 2,000 rules of the Aṣṭādhyāyī.