Supernova-NepaliFast-V4 / final_supernova_vs_nepalibpe_benchmark_report.md
Supernova11c's picture
Upload 2 files
c67432e verified
|
Raw
History Blame Contribute Delete
9.57 kB

Supernova vs NepaliBPE

Final Forensic Tokenizer Benchmark Report

Report generated: 2026-09-04 02:17:25


1. Executive Verdict

Overall winner: Supernova

The comparison was performed using two independent dimensions:

  1. Runtime performance
  2. Tokenizer quality / forensic correctness

Runtime speed was not used as a substitute for tokenizer quality.

The quality benchmark was treated as the primary indicator of tokenizer strength because tokenization quality directly affects text representation, information preservation, Unicode handling, unknown-token behavior, and downstream language-model usability.


2. Benchmark Scope

Supernova

Repository:

Supernova11c/Supernova-NepaliFast-V4

NepaliBPE

Repository:

Aananda-giri/NepaliBPE

The two tokenizers were tested against the same test inputs.

The benchmark included:

  • Nepali
  • English
  • mixed Nepali/English
  • Unicode
  • Devanagari
  • combining characters
  • emoji
  • numbers
  • dates
  • punctuation
  • whitespace
  • rare Nepali vocabulary
  • adversarial words
  • long documents
  • reconstruction
  • unknown-token detection
  • grapheme preservation
  • morphological consistency
  • determinism
  • token efficiency

3. Runtime Performance

The previously completed large-scale speed benchmark used:

  • Documents: 420,000
  • Characters: 9,690,000
Metric Supernova NepaliBPE
Tokens 6,710,000 4,580,000
Mean time 10.0207 s 32.0435 s
Best time 5.6490 s 29.5526 s
Characters/sec 967,002 302,401
Tokens/sec 669,617 142,931

Speed result

Supernova wins the runtime benchmark.

Supernova processed characters approximately 3.20× faster.

Supernova processed emitted tokens approximately 4.68× faster.

However, this does not by itself establish that Supernova is the better tokenizer.


4. Quality / Forensic Benchmark

Unlike the runtime test, this benchmark evaluates what the tokenizer actually does with text.

Quality comparison

Metric Supernova NepaliBPE Winner
Exact reconstruction 81.25% 34.03% Supernova
NFC-equivalent reconstruction 81.25% 34.03% Supernova
Unknown-token rate 0.03% 50.15% Supernova
Character coverage 99.98% 78.32% Supernova
Grapheme preservation 96.47% 68.47% Supernova
Token efficiency 1.3569 3.9502 NepaliBPE
Nepali fragmentation 6.3875 1.3034 NepaliBPE
Morphological consistency 100.00% 100.00% Tie
Determinism 100.00% 100.00% Tie

5. Quality Win Count

Wins
Supernova 5
NepaliBPE 2
Ties 2
Unknown 0

Quality winner

Supernova

This result is based on independent benchmark dimensions rather than tokenizer ID comparisons.


6. Why Token IDs Were Not Compared

Token IDs from two independent tokenizers do not have equivalent meaning.

For example:

  • Supernova token ID 100
  • NepaliBPE token ID 100

does not mean that those tokens represent the same linguistic unit.

Therefore this benchmark compares:

  • emitted token text
  • reconstruction
  • unknown-token strings
  • Unicode preservation
  • text boundaries
  • morphological behavior

rather than numerical token IDs.


7. Unknown-Token Methodology

Unknown-token detection was performed against the actual emitted token strings.

Recognized unknown representations included forms such as:

  • [UNK]
  • <unk>
  • <UNK>
  • <|unk|>
  • <unknown>

This avoids relying exclusively on a tokenizer API's unk_token_id.

Therefore, an unknown token is counted when the tokenizer actually emits an unknown representation.


8. Reconstruction

Exact reconstruction tests whether:

decode(encode(text)) == text

NFC-equivalent reconstruction separately tests Unicode canonical equivalence.

These are deliberately separate measurements.

A tokenizer that inserts punctuation spacing, changes characters, or loses unknown characters should not receive an exact-reconstruction score merely because the resulting text looks approximately similar.


9. Unicode and Grapheme Testing

The benchmark tested:

  • Devanagari
  • combining marks
  • Latin accented characters
  • multiple writing systems
  • mathematical symbols
  • currency symbols
  • emoji
  • complex Unicode sequences

Grapheme preservation was also tested because a Unicode character sequence can represent a single user-perceived character.

This is especially important for a tokenizer intended to support Nepali and multilingual text.


10. Nepali Quality

Nepali quality was evaluated separately from generic token efficiency.

The benchmark examined:

  • Nepali word fragmentation
  • complex Devanagari sequences
  • conjuncts
  • vowel signs
  • rare Nepali words
  • related-word consistency
  • Nepali/English mixing

The benchmark does not assume that fewer tokens automatically means better linguistic quality.


11. Morphological Consistency

Morphological consistency was calculated using the actual textual representation of tokens.

Token IDs were intentionally excluded.

Related word families such as:

  • नेपाल / नेपालमा
  • नेपाली / नेपालीहरू
  • शिक्षा / शिक्षाले
  • विज्ञान / वैज्ञानिक
  • समाज / सामाजिक
  • राष्ट्र / राष्ट्रिय
  • स्वतन्त्र / स्वतन्त्रता

were examined for preservation of shared textual structure.


12. Token Efficiency

Token efficiency is reported independently as:

characters per token

Higher values generally indicate that more input text is represented per token.

However:

Token efficiency is NOT equivalent to linguistic quality.

A tokenizer can achieve fewer tokens by aggressively merging text without necessarily producing a better representation.

Therefore this metric was never allowed to determine the final winner by itself.


13. Determinism

Both tokenizers were tested repeatedly on identical inputs.

A deterministic tokenizer should produce the same token sequence every time.

This is important for reproducibility and model training.


14. Final Assessment

Supernova strengths

Based on the completed benchmarks, Supernova's strongest areas are the dimensions in which it won above.

In particular, its runtime performance was:

3.20× the character throughput of NepaliBPE

and:

4.68× the token throughput of NepaliBPE

The quality benchmark provides the independent evidence for whether that performance is accompanied by strong text representation.

NepaliBPE strengths

NepaliBPE's strongest areas are likewise the dimensions in which it won above.

Its 50,006-token vocabulary and BPE representation can provide substantially different token segmentation from Supernova's custom longest-match approach.

The final quality table—not popularity, search ranking, downloads, or vocabulary size—is what determines the technical comparison.


15. Speed vs Quality

The results should be interpreted as two separate questions.

Question A — Which tokenizer is faster?

Supernova

Question B — Which tokenizer demonstrated stronger quality across the tested dimensions?

Supernova

Question C — Which tokenizer is the stronger overall candidate?

Supernova

The overall verdict prioritizes tokenizer quality rather than allowing raw speed to compensate for information loss or unknown-token failures.


16. Important Limitations

This is a serious engineering benchmark, but it is not equivalent to training a large language model with each tokenizer and measuring downstream perplexity.

A definitive production-level tokenizer evaluation would additionally include:

  • large real-world Nepali corpora
  • held-out Nepali corpora
  • downstream language-model training
  • validation perplexity
  • compression ratio
  • vocabulary utilization
  • fertility
  • sequence-length distribution
  • embedding/model compatibility
  • multilingual robustness
  • domain-specific corpora
  • OCR/noisy-text robustness

Therefore the conclusion here means:

Winner under the tested tokenizer-level benchmark conditions.

It does not claim that one tokenizer is universally superior for every possible language-model architecture.


17. Final Verdict

🏆 Supernova

The benchmark evidence indicates that Supernova is the stronger overall tokenizer under the completed test suite.

The conclusion is based on measured behavior rather than:

  • Google ranking
  • popularity
  • repository fame
  • vocabulary size
  • download count
  • marketing claims
  • arbitrary speed weighting

The final decision comes from reproducible benchmark measurements.


18. Benchmark Integrity Statement

This comparison deliberately avoids several common benchmarking errors:

  • No cross-tokenizer token-ID comparison
  • No incorrect unk_token_id assumption
  • No speed-only winner declaration
  • No automatic assumption that fewer tokens means better language quality
  • No arbitrary claim that vocabulary size equals tokenizer quality
  • No hidden competitor-specific advantage
  • Same corpus categories for both tokenizers
  • Fresh tokenizer loading
  • Raw token evidence inspected
  • Unknown tokens counted from emitted token strings
  • Reconstruction tested independently

FINAL RESULT

Speed winner: Supernova

Quality winner: Supernova

Overall winner: Supernova

Benchmark status: COMPLETE