SampoNLP: A Self-Referential Toolkit for Morphological Analysis of Subword Tokenizers Iaroslav Chelombitko, Ekaterina Chelombitko, Aleksey Komissarov Abstract The quality of subword tokenization is critical for Large Language Models, yet evaluating tokenizers for morphologically rich Uralic languages is hampered by the lack of clean morpheme lexicons. We introduce SampoNLP, a corpus-free toolkit for morphological lexicon creation using MDL-inspired Self-Referential Atomicity Scoring, which filters composite forms through internal structural cues - suited for low-resource settings. Using the high-purity lexicons generated by SampoNLP for Finnish, Hungarian, and Estonian, we conduct a systematic evaluation of BPE tokenizers across a range of vocabulary sizes (8k–256k). We propose a unified metric, the Integrated Performance Score (IPS), to navigate the trade-off between morpheme coverage and over-splitting. By analyzing the IPS curves, we identify the "elbow points" of diminishing returns and provide the first empirically grounded recommendations for optimal vocabulary sizes (k) in these languages. Our study not only offers practical guidance but also quantitatively demonstrates the limitations of standard BPE for highly agglutinative languages. The SampoNLP library and all generated resources are made publicly available 1 1 1 https://github.com/AragonerUA/SampoNLP . 1 Introduction The performance of subword tokenization algorithms like Byte-Pair Encoding (BPE) 21 is a cornerstone of modern Natural Language Processing (NLP). While highly effective for many languages, their purely statistical nature poses a significant challenge for morphologically rich, agglutinative languages 3 ; 19 . In the Uralic family, a group of languages known for its complex morphology and diverse linguistic phenomena 10 , words are often long concatenations of morphemes (e.g., Finnish talo-i-ssa-ni-ko-kaan - "not in my houses either?"). For such languages, the quality of tokenization is not just an engineering detail but a critical factor that determines a model’s ability to grasp grammatical structure and generalize effectively 9 ; 8 . This raises a pressing, yet under-explored, practical question, known to be a challenge in Uralic NLP: What is the optimal tokenizer vocabulary size (k) to achieve robust morphological representation? The importance of this question was highlighted by recent work demonstrating the benefits of specialized tokenizers for these languages 4 . Addressing this question reveals a more fundamental problem: the scarcity of high-purity morphological resources for evaluation. While lexical data is available in spell-checking dictionaries, their raw combination of stems and affixes results in a noisy candidate list. Manual curation is not scalable, and established corpus-based methods like Morfessor 7 are ill-suited for the many low-resource Uralic languages 1 . To address this challenge, we present SampoNLP, a toolkit based on a corpus-free and self-referential pipeline for refining morphological lexicons. The proposed method, "MDL-inspired Self-Referential Atomicity Scoring," draws its theoretical motivation from the Minimum Description Length principle 18 , but adapts it to a type-only setting. The core algorithm iteratively estimates the atomicity of each candidate, distinguishing between simple and composite forms by analyzing internal structural patterns within the dataset itself. This lightweight and reproducible approach offers a practical way to produce cleaner morphological resources, a recognized need for data-scarce environments where traditional corpus-based methods are not viable 10 . Having established a robust methodology for resource creation, we leverage our generated lexicons to address the core problem of this paper: the vocabulary-morphology trade-off inherent in BPE tokenization 3 . We conducted a systematic evaluation of BPE tokenizers for Finnish, Hungarian, and Estonian across vocabulary sizes from 8k to 256k. The development of novel evaluation frameworks that go beyond downstream performance is a growing area of research 5 . In line with this, to precisely navigate the aforementioned trade-off, we introduce the Integrated Performance Score (IPS), a single metric that balances Lexical Morpheme Coverage (LMC) against the Over-Split Rate (OSR). This allows us to model the performance curve and identify the optimal vocabulary range, providing a principled answer to our central research question. Our contributions are thus twofold and equally significant: 1. A Corpus-Free Morphological Method: We introduce a fully automatic and reproducible pipeline for refining morphological lexicons without relying on corpus frequencies or external resources, released as an open-source toolkit, SampoNLP . 2. A Quantitative Evaluation: We conduct a systematic analysis of BPE tokenizers for Finnish, Estonian, and Hungarian, examining how vocabulary size affects morphological granularity through newly defined metrics of coverage and over-segmentation. 2 Related Work The evaluation and optimization of subword tokenization for morphologically rich languages intersects several research areas: subword tokenization algorithms, unsupervised morphological analysis, rule-based analyzers, and language-specific NLP for Uralic languages. 2.1 Subword Tokenization and Morphology Byte-Pair Encoding (BPE) 21 has become the de facto standard for subword tokenization in modern NLP. Alongside it, methods like the Unigram Language Model 13 have been proposed, but the purely statistical nature of these approaches presents well-documented challenges for morphologically complex languages. The work of 3 demonstrated that BPE tokenizers often fail to align with linguistic morpheme boundaries. Interestingly, parallel challenges in identifying meaningful subsequence units have been explored in domains beyond NLP, such as the tokenization of biological sequences like primate genomes 17 . The question of optimal vocabulary size has often been guided by heuristics or evaluated indirectly via downstream task performance 14 . Our work directly addresses this gap by proposing a methodology for intrinsic, morphologically-grounded evaluation to provide data-driven recommendations for Uralic languages. 2.2 Unsupervised Morphological Analysis The unsupervised discovery of morphological structure has a rich history. One major family of approaches relies on statistical cues from corpora to identify boundaries. Classic methods such as Branching Entropy and Accessor Variety 6 analyze the predictability of subsequent characters to hypothesize morpheme breaks. Another prominent family of methods is based on the Minimum Description Length (MDL) principle. Morfessor 7 and its variants represent the canonical probabilistic approach, finding a lexicon that best compresses a text corpus. While successful, these methods are fundamentally corpus-based, requiring token frequency information that may not be available in low-resource settings. Our approach, while MDL-inspired, operates in a corpus-free, type-only regime. It represents a different paradigm: self-referential filtering of a candidate list. By operating purely on the internal structure of a candidate set, we provide a lightweight method suited to resource-scarce scenarios, a persistent challenge in Uralic NLP 1 . 2.3 Rule-Based Analyzers and Tokenization for Uralic Languages For Uralic languages, rule-based morphological analyzers built on Finite-State Transducers (FSTs) like Omorfi 16 and the GiellaLT 2 2 2 https://giellalt.github.io/ infrastructure 11 are invaluable resources. While their generative outputs are linguistically comprehensive, they are not directly optimized for use as a minimal reference morphemes lexicon. Our IMDP pipeline offers a contrasting approach: a data-driven methodology for distilling such a lexicon from a type-only candidate list, as can be extracted from dictionary-based resources like Hunspell, without requiring token frequencies from a corpus. The challenge of effective tokenization for this language family has recently gained significant attention. Broader findings have established that language-specific modeling is crucial for morphologically rich languages, with studies on Finnish demonstrating clear benefits of monolingual models like FinBERT over multilingual ones 22 . Building on this principle, a recent study by 4 specifically addressed the severe under-representation of Uralic languages in large multilingual models. They demonstrated that training specialized, large-vocabulary monolingual tokenizers yields substantial improvements in compression efficiency. However, while establishing the need for specialized resources, their work left the question of how to determine an optimal vocabulary size open for future investigation. Concurrently, the need for better evaluation metrics has become a prominent research topic. The Qtok framework 5 , for instance, proposed a comprehensive approach to evaluating multilingual tokenizer quality, while other studies have also advocated for moving beyond downstream task performance towards more intrinsic, linguistically-informed measures 2 . Our Integrated Performance Score (IPS) directly addresses this call from the community for more morphologically-grounded metrics. Our current work builds on these foundations. It utilizes similar high-quality data sources as those in 4 to train the tokenizers being evaluated. Furthermore, by proposing a concrete methodology, it answers the call for better evaluation and finds the optimal vocabulary sizes that the former study alluded to, thus providing a logical next step in this line of research. 3 Methodology. The IMDP Pipeline Figure 1: An overview of the Iterative Morphological Decomposition Pipeline (IMDP). To create a high-purity morpheme lexicon from a noisy, raw list of candidate forms, we propose the Iterative Morphological Decomposition Pipeline (IMDP). Our approach is designed to be fully automatic and operates in a corpus-free, type-only regime, requiring only the candidate list as input. The core of the pipeline is a method we term "MDL-inspired Self-Referential Atomicity Scoring," which iteratively evaluates how "fundamental" each candidate is relative to the entire set. The entire process is visualized in Figure 1 . The pipeline consists of three main stages: (1) Pre-filtering and Initial Scoring, (2) Iterative Score Refinement, and (3) Final Filtering via Automated Thresholding. 3.1 Stage 1: Candidate Pre-filtering and Initial Scoring This initial stage aims to drastically reduce non-linguistic noise and establish a baseline score for each plausible candidate. 3.1.1 Hard Pre-filtering First, we apply a series of deterministic filters to the raw input list C_{raw} . A token t\in C_{raw} is discarded if it: 1. Contains symbols from a non-target script (e.g., Cyrillic in a Latin-based list). We define a valid character set \Sigma for each language (e.g., [a-záéíóöőúüű] for Hungarian). 2. Contains any non-alphabetic characters (e.g., numbers, punctuation, URLs), excluding initial/final hyphens used to mark affixes. 3. Is a proper noun or acronym (heuristic: starts with a capital letter or consists of multiple uppercase letters). 4. Is excessively long ( |t|>30 ) or too short ( |t|S_{k}(t).\end{cases} This update rule creates a competitive dynamic where atomic morphemes retain high scores, while composite words are iteratively penalized towards zero. 3.2.3 Convergence The iterative process continues until the system reaches a stable state. We define convergence as the point where the maximum absolute change in any token’s score between two consecutive iterations falls below a small threshold \max_{t\in\mathcal{C}}\bigl|S_{k+1}(t)-S_{k}(t)\bigr|<\varepsilon We use \varepsilon=1e-7 and a safeguard limit of max\_iterations=100 . 3.3 Stage 3: Final Filtering via Automated Thresholding After the scores converge, the final distribution of scores typically shows a heavy concentration of composite candidates at very low scores, while atomic candidates retain higher scores. To automatically and reproducibly determine a separation threshold between these groups, we employ Otsu’s method 15 . Originally developed for image processing to separate foreground from background, this algorithm finds an optimal threshold \tau for a distribution by maximizing the inter-class variance between the two resulting classes (in our case, "atomic" vs. "composite"). This data-driven approach avoids manual parameter tuning and adapts to the specific score distribution of each dataset. All tokens t with a final score S_{final}(t)>=\tau are classified as atomic and form our final, high-purity morpheme lexicon. Lang Initial Atomic Reduct Reduct Cands Morphs % Factor Fin 499,647 3,850 99.23% 129.8x Est 281,256 5,705 97.97% 49.3x Hung 103,317 3,189 96.91% 32.4x Table 1: Efficiency of the IMDP pipeline in cleaning and reducing morpheme candidate lists. 4 Experimental Setup To evaluate the impact of vocabulary size on morphological coverage, we conducted a systematic analysis for three Uralic languages: Finnish, Hungarian, and Estonian. Our experimental setup consists of three main stages: creating the reference morphemes, training the tokenizers, and defining the evaluation metrics. 4.1 Data Our methodology requires two types of data for each language: a raw list of morpheme candidates for cleaning and a large text corpus for tokenizer training. 1. Morpheme Candidate Lists: The initial "dirty" lists of candidates were constructed from authoritative, open-source spell-checking dictionaries based on the Hunspell framework 3 3 3 https://hunspell.github.io/ . For Hungarian and Estonian, we utilized the comprehensive dictionaries curated by The LibreOffice Project 4 4 4 https://github.com/LibreOffice/dictionaries . For Finnish, which requires special handling of compounds, we used the dedicated dictionary from the hunspell-fi project 5 5 5 https://github.com/fginter/hunspell-fi . For each language, the full set of unique stems (from.dicfiles) and affixes (from.afffiles) was merged to create a comprehensive but structurally noisy candidate list, which serves as the input to our IMDP pipeline. This approach of leveraging widely available dictionary resources provides a practical starting point for morphological analysis. 2. Text Corpora: For training the BPE tokenizers, we used large, pre-processed corpora derived from Wikipedia snapshots 6 6 6 https://dumps.wikimedia.org . Our choice of data source and preprocessing methodology aligns with previous work on creating specialized Uralic tokenizers 4 , ensuring a comparable basis for our analysis. It is critical to emphasize that these corpora were used exclusively for training the BPE tokenizers and were not used in any stage of our morpheme list refinement pipeline, thus preserving the corpus-free nature of the IMDP method. 4.2 Reference Lexicon Creation For each of the three languages, we applied our Iterative Morphological Decomposition Pipeline (IMDP), as described in Section 3, to the corresponding raw candidate list. The pipeline was configured with the following parameters: a minimum morpheme length min\_length=1 , a minimum type-support m=3 , and a convergence threshold \varepsilon=1e-7 . The process was run until convergence. The final filtering was performed using the automatically determined Otsu threshold 15 . This procedure yielded three high-purity reference morpheme lexicons ( G_{fin} , G_{hun} , G_{est} ), the statistics of which are summarized in Table 1 . 4.3 Tokenizer Training Using the tokenizers library 7 7 7 https://github.com/huggingface/tokenizers and SentencePiece 12 for comparison, we trained a series of Byte-Pair Encoding (BPE) tokenizers for each language from scratch. The tokenizers were trained on the respective Wikipedia corpora. To analyze the effect of vocabulary size, we trained separate models for a range of vocabulary sizes k, starting from 8,000 and up to 256,000 ( k\in {8k, 16k, 32k, 40k, 50k, 64k, 80k, 100k, 128k, 150k, 180k, 200k, 220k, 240k, 256k}). All tokenizers were trained with a min\_frequency of 2 for merges. 4.4 Evaluation Metrics To provide a nuanced and rigorous evaluation of tokenizer quality, we must account for the fundamental trade-off between morphological coverage and over-segmentation. A tokenizer that perfectly represents all morphemes (high coverage) but also excessively splits common words is not optimal. To capture this balance in a single, unified score, we introduce the Integrated Performance Score (IPS). The IPS models this trade-off geometrically. We consider a 2D space where the ideal tokenizer resides at the point (Coverage=1, OverSplit=0). The IPS of any real tokenizer is its normalized Euclidean distance from this ideal point, scaled to a [0, 1] range where 1 is perfect. First, we define the two core components: 1. Lexical Morpheme Coverage (LMC): The fraction of atomic morphemes from our reference lexicon G that are perfectly represented as a single token in the tokenizer’s vocabulary V_{k} . This measures the tokenizer’s lexical "knowledge" of fundamental morphological units. \mathrm{LMC}\;=\;\frac{\bigl|\{\,m\in G\;\mid\;m\in V_{k}\,\}\bigr|}{|G|}. 2. Over-split Rate (OSR): The fraction of morphemes from G that the tokenizer fails to represent as single tokens, thus always splitting them into multiple pieces. \mathrm{OSR}=\frac{\left|\left\{\,m\in M\,\middle|\,\begin{subarray}{c}m\text{ occurs in $\geq$1 word}\\ m\text{ never as a single token}\end{subarray}\right\}\right|}{\left|\left\{\,m\in M\,\middle|\,m\text{ in $\geq$1 word}\right\}\right|}. From these, the Integrated Performance Score (IPS) is calculated as: IPS=1-(\frac{\sqrt{(1-LMC)^{2}+OSR^{2}}}{\sqrt{2}}) This single metric allows for a clear and direct comparison of tokenizers across different vocabulary sizes. A higher IPS indicates a better balance between representing morphemes and avoiding excessive fragmentation. Our final analysis of optimal vocabulary sizes is based on identifying the "elbow point" on the IPS vs. vocabulary size curve. 5 Results and Analysis Our experiment yielded clear and significant patterns regarding the relationship between tokenizer vocabulary size and morphological performance. To capture the fundamental trade-off between coverage and over-segmentation, we analyzed the Integrated Performance Score (IPS) for each language. The resulting IPS curves for Estonian (Figure 5 ), Finnish (Figure 6 ), and Hungarian (Figure 4 ) clearly show the performance profile for each language. Supplementary details on the component metrics (LMC and OSR) available in Figures 2 and 3 . Figure 2: Lexical Morpheme Coverage (LMC) across different vocabulary sizes (k). LMC represents the percentage of reference morphemes found as single, complete tokens in the tokenizer’s vocabulary. Figure 3: Over-Split Rate (OSR) as a function of vocabulary size (k). OSR denotes the fraction of reference morphemes that occur in words but never appear as a single token in any tokenization. 5.1 General Observation: A Clear Trade-off Profile The IPS curves for all three languages exhibit a classic logarithmic growth pattern, demonstrating the law of diminishing returns. The score increases rapidly for smaller vocabulary sizes, indicating that initial additions to the vocabulary are highly efficient at capturing morphological structure. However, the rate of improvement progressively slows, showing that ever-larger vocabularies provide only marginal gains at a significant cost to model size. This confirms that a "sweet spot" or an optimal range exists for each language. 5.2 Cross-Linguistic Analysis: Three Distinct Performance Tiers The results reveal three distinct performance tiers, highlighting the varying degrees to which standard BPE can model the morphology of these languages. 1. Hungarian (hu): As shown in Figure 4 , Hungarian demonstrates by far the best performance. Its IPS curve starts at 0.29 and rises sharply, reaching a maximum of 0.73. This high score suggests that BPE is reasonably effective at learning the statistical regularities of Hungarian morphology. 2. Estonian (et): Estonian occupies the middle tier, with its IPS curve depicted in Figure 5 . The score starts at 0.22 and reaches a maximum of 0.39. While better than Finnish, this score indicates that less than 40% of the "ideal" tokenizer performance is achieved, even with a large vocabulary. 3. Finnish (fi): Figure 6 illustrates the most challenging profile for Finnish. With a maximum IPS of only 0.31, the results quantitatively demonstrate that standard BPE is fundamentally ill-suited for capturing the complexities of Finnish morphology. Figure 4: IPS vs. vocabulary size (k) for Hungarian . Hungarian shows the most consistent improvement in IPS, reflecting its comparatively transparent agglutinative structure with fewer morphophonological alternations. The elbow point is at 80k, and the 90% quality threshold at 128k, yielding a recommended range of 80k–128k. 5.3 Identifying the Optimal Vocabulary Range (k*) To determine a practical and effective vocabulary size, we define a recommended range for k*. The lower bound of this range is the "elbow" point (k_elbow), identified by the Kneedle algorithm 20 , which marks the point of diminishing returns. The upper bound is the 90% quality point (k_q90), where 90% of the maximum observed IPS is achieved. As shown in Figures 4 , 6 , 5 , and summarized in Table 2 , this analysis leads to the following recommendations: Lang Max Gain Point (k_gain) Elbow Point (k_elbow) 90% Quality Point (k_q90) Recommend k* Range Hung 40,000 80,000 128,000 80k – 128k Est 16,000 80,000 128,000 80k – 128k Fin 64,000 80,000 150,000 80k – 150k Table 2: Key points on the IPS curve for determining the optimal vocabulary range. 1. Hungarian (hu): The IPS curve for Hungarian (Figure 4 ) shows a clear optimal range between k=80,000 and k=128,000. The elbow is found at 80k, and 90% of the maximum performance is reached at 128k. As visualized on the plot, expanding the vocabulary beyond this range yields only minimal performance gains. 2. Estonian (et): For Estonian (Figure 5 ), the recommended range is also k=80,000 to k=128,000. Similar to Hungarian, the elbow is at 80k and the 90% quality mark is at 128k, establishing this as the zone of best compromise between performance and size. 3. Finnish (fi): The analysis for Finnish (Figure 6 ) indicates a need for a larger vocabulary. The elbow is at k=80,000, but to achieve 90% of the (albeit low) maximum performance, a vocabulary of k=150,000 is required. This suggests that for Finnish, the optimal range is k=80,000 to k=150,000, reflecting the language’s high morphological complexity. Figure 5: IPS vs. vocabulary size (k) for Estonian . While the overall pattern of diminishing returns is similar to Hungarian, the lower IPS plateau indicates reduced learnability due to Estonian’s extensive morphophonological alternations, which obscure orthographic morpheme boundaries. The recommended range remains 80k–128k. These findings provide a quantitative foundation for the critical decision of vocabulary sizing, transforming it from a heuristic-based choice into a principled optimization problem. Complete numerical results for all evaluated vocabulary sizes are provided in Appendix A (Table 3 ) for reference. 6 Conclusion In this work, we addressed the dual challenge of creating high-purity morphological resources in a corpus-free setting and using them to evaluate subword tokenizers for Uralic languages. We introduced SampoNLP, a toolkit featuring a novel pipeline based on "MDL-inspired Self-Referential Atomicity Scoring," which successfully refines noisy candidate lists into clean morpheme lexicons. Applying these lexicons, our systematic evaluation of BPE tokenizers yielded two key findings. First, we provide an empirically-grounded recommendations for optimal vocabulary sizes, identifying a range of 80k-128k for Hungarian and Estonian, and 80k-150k for Finnish, as the most effective trade-off between performance and model size. Second, our results quantitatively demonstrate the severe limitations of standard BPE for highly agglutinative languages like Finnish, where performance plateaus at a strikingly low level. This study confirms that while vocabulary size optimization is a crucial step, it is not a panacea. We release our SampoNLP library and the generated morpheme lists to the community to facilitate reproducible research and encourage the development of more morphologically-aware tokenization methods for the Uralic language family. Figure 6: IPS vs. vocabulary size (k) for Finnish . Finnish exhibits the lowest IPS plateau, consistent with its rich system of consonant gradation and stem alternations, which make orthographic segmentation less stable for BPE. The elbow is at 80k, while 90% of the maximum IPS is reached at 150k, suggesting a recommended range of 80k–150k. Discussion Our results yield two key insights. First, the effectiveness of BPE varies dramatically by language: while Hungarian achieves a high IPS (max \sim 0.73), the low scores for Finnish ( \sim 0.31) and Estonian ( \sim 0.39) quantitatively demonstrate the algorithm’s fundamental limitations for these highly agglutinative languages. Second, for all languages, an empirically identifiable "sweet spot" for vocabulary size exists, beyond which performance gains diminish. Here, “optimality” is understood as morphological sufficiency - the point at which the tokenizer captures the productive structure of a language with minimal redundancy. This notion is intrinsic by design, offering a language-level criterion rather than task-specific optimization. We acknowledge the limitations of our approach. The IPS metric abstracts away qualitative segmentation differences - a necessary compromise for scalability. Our use of clean, standardized corpora also isolates the variable of vocabulary size but does not reflect the noise of real-world data. These aspects represent clear avenues for future work. While our method produces a refined set of recurrent sub-lexical units, we do not claim full linguistic morpheme correctness. The IMDP segmentation is orthographic and self-referential in nature, providing a practical approximation rather than a phonologically grounded morphological analysis. In conclusion, our findings suggest that while optimizing k* is a crucial step, it may be insufficient for languages like Finnish. The low performance ceiling for BPE underscores the need for morphologically-aware tokenization methods. We believe our SampoNLP toolkit and the generated lexicons provide the community with a reproducible benchmark to develop and test such new strategies. Appendix A Appendix Language Vocabulary Size (k) Total Morphemes Morpheme Coverage % Over-Split Rate % Estonian 8,000 5,705 11.27% 65.79% Estonian 16,000 5,705 13.71% 59.13% Estonian 32,000 5,705 16.49% 53.65% Estonian 40,000 5,705 17.48% 51.98% Estonian 50,000 5,705 18.18% 51.43% Estonian 64,000 5,705 19.30% 50.32% Estonian 80,000 5,705 20.68% 49.13% Estonian 100,000 5,705 21.74% 48.49% Estonian 128,000 5,705 22.99% 47.86% Estonian 150,000 5,705 23.79% 47.46% Estonian 180,000 5,705 24.78% 47.06% Estonian 200,000 5,705 25.38% 46.43% Estonian 220,000 5,705 25.74% 46.19% Estonian 240,000 5,705 26.23% 46.19% Estonian 256,000 5,705 26.81% 46.11% Finnish 8,000 3,850 7.85% 78.96% Finnish 16,000 3,850 9.76% 74.37% Finnish 32,000 3,850 12.20% 69.13% Finnish 40,000 3,850 12.95% 68.04% Finnish 50,000 3,850 13.95% 66.62% Finnish 64,000 3,850 15.53% 64.73% Finnish 80,000 3,850 16.63% 63.36% Finnish 100,000 3,850 17.84% 62.46% Finnish 128,000 3,850 19.11% 61.61% Finnish 150,000 3,850 20.21% 61.18% Finnish 180,000 3,850 21.51% 60.52% Finnish 200,000 3,850 22.16% 60.28% Finnish 220,000 3,850 22.93% 60.05% Finnish 240,000 3,850 23.38% 59.95% Finnish 256,000 3,850 23.73% 59.81% Hungarian 8,000 3,189 25.15% 67.72% Hungarian 16,000 3,189 34.34% 57.01% Hungarian 32,000 3,189 45.03% 46.14% Hungarian 40,000 3,189 49.23% 42.17% Hungarian 50,000 3,189 52.46% 39.97% Hungarian 64,000 3,189 56.98% 37.84% Hungarian 80,000 3,189 60.90% 35.72% Hungarian 100,000 3,189 64.88% 34.33% Hungarian 128,000 3,189 69.24% 33.27% Hungarian 150,000 3,189 71.97% 32.37% Hungarian 180,000 3,189 74.29% 32.28% Hungarian 200,000 3,189 75.67% 32.16% Hungarian 220,000 3,189 77.08% 32.00% Hungarian 240,000 3,189 78.43% 31.92% Hungarian 256,000 3,189 79.12% 32.04% Table 3: Detailed experimental results for BPE tokenizers of varying vocabulary sizes across three Uralic languages. Morpheme Coverage represents the percentage of reference morphemes found in the vocabulary (LMC). Over-Split Rate is the percentage of reference morphemes with support in W that never appear as a single token in any tokenization.