Sanskrit stemmer benchmarks
Benchmark-only dictionary — not distributed
san-default is privately retained for reproducible measurement only. Radixor does not distribute, install, publish, register, or support this dictionary because file-applicable redistribution rights are unresolved. It has no Maven coordinates, BOM/catalog entry, standard-package membership, or Python package.
Evidence: README licenses only alternate san2 data, not the selected san file. Benchmarking these private bytes does not assert a right to redistribute them.
Dictionary size: . The exact count is 55 distinct usable word forms after parser-compatible filtering and exact case-preserved deduplication. Stars rank size among all 167 benchmarked dictionaries; they do not measure linguistic quality, availability, or benchmark accuracy.
The corpus, quality, generalization, and speed evidence below belongs to the private 2026-09-11 Radixor/Java 4.4.0 documentation snapshot. It is never placed on a production or release classpath.
Dictionary corpus
| Model ID | Private version | Language | Dictionary rows | Distinct usable forms | Complete tokens | Already-root | Changed | Timing workload | Timing tokens |
|---|---|---|---|---|---|---|---|---|---|
san-default |
1.0.0 |
SAN |
54 | 55 | 109 | 108 | 1 | changed tokens | 5,000 |
Radixor patch-command distribution
| Command class | Complete-corpus forms |
|---|---|
AppendCharacterCommand |
1 |
PreserveCommand |
108 |
Exact-root accuracy
| Stemmer | All exact | Changed exact | Root preserved |
|---|---|---|---|
| Private Radixor filesystem model | 100.000% | 100.000% | 100.000% |
Pairwise linguistic quality
OI and UI report over-stemming and under-stemming with raw numerators and denominators. A zero denominator is n/a.
| Stemmer | Dictionary mode | OI raw | OI | UI raw | UI |
|---|---|---|---|---|---|
| Private Radixor filesystem model | ALL_WORDS |
0 / 1484 | 0.000000% | 0 / 1 | 0.000000% |
| Private Radixor filesystem model | LOWERCASE_GROUPS_ONLY |
0 / 1484 | 0.000000% | 0 / 1 | 0.000000% |
Dictionary-family generalization
This frozen five-seed curve measures transfer to withheld lexical families; it does not establish redistribution or support status.
| Training rows | Median unseen occurrences | Unseen all exact | Unseen changed exact | Unseen root preserved |
|---|---|---|---|---|
| 10% | 99 | 98.990% | 0.000% | 100.000% |
| 20% | 87 | 98.851% | 0.000% | 100.000% |
| 30% | 77 | 98.701% | 0.000% | 100.000% |
| 40% | 65 | 98.462% | 0.000% | 100.000% |
| 50% | 55 | 98.182% | 0.000% | 100.000% |
| 60% | 45 | 97.778% | 0.000% | 100.000% |
| 70% | 32 | 100.000% | 0.000% | 100.000% |
| 80% | 22 | 100.000% | n/a | 100.000% |
| 90% | 10 | 100.000% | n/a | 100.000% |
| 100% | 0 | n/a | n/a | n/a |
Runtime timing
JMH average time uses 3 one-second warmup iterations, 5 one-second measurement iterations, 3 independent forks, and 1 thread.
| Benchmark | Score ms/op | Error ms | ns/token |
|---|---|---|---|
| Private Radixor filesystem model | 0.078 | 0.000 | 15.6 |