Toppling the Hierarchy in Byte-level Language Modeling

Abstract

This work examines recent byte-level models and their failure to perfectly manipulate characters. State-of-the-art byte-level models use a hierarchical structure, starting at the byte level, downsampling to the word level, and then upsampling back to bytes. While this improves training and inference efficiency, we find that the hierarchical design itself limits character-level understanding, with pure byte-level models consistently outperforming hierarchical variants on character manipulation tasks. Ablating transformer layers into attention and feed-forward components further reveals that byte-level attention is the primary mechanism driving this behavior. Together, our results provide an explanation for the character-level failures of hierarchical byte models and establish a clear trade-off between computational efficiency and fine-grained character understanding.

1 Introduction

Byte-level or character-level models have shown promise in recent years, with several works boasting superior performance over BPE-based subword models on character-level manipulation benchmarks (i.e., CUTE Edman et al. (2024) and the multilingual version EXECUTE Edman et al. (2025), while keeping up-to-par on other popular benchmarks (i.e., ARC Clark et al. (2018), MMLU Hendrycks et al. (2021), PiQA Bisk et al. (2019), inter alia).

However, despite the performance being superior to subword models, it is still far from perfect. The Byte Latent Transformer Pagnoni et al. (2025), which is pretrained from scratch, only achieved an average of 54.1% on CUTE, and the more recent Bolmo model Minixhofer et al. (2026), which is adapted from Olmo, only increased CUTE performance from its equivalent subword-level counterpart by around 6% (78.6% vs. 72.9%). Meanwhile, Edman et al. (2025) found that even BPE-based LLMs can achieve remarkable, near 100%, few-shot EXECUTE scores in Amharic and other low-resource languages with almost no pre-training data. One reason for this stellar performance may be that these LLMs typically dedicate no BPE merges to the Amharic script and other minority scripts, so with byte fallback, they end up operating entirely on the byte-level, making character-level manipulation easier.

Figure 1: Hierarchies tested in this work, with word-level transformer layers in the middle (a.k.a., hourglass), on the bottom, top, and interleaved. Dashed lines indicate skip connections.

These findings raise doubts about the dominant hierarchical approach to byte-level modeling, where most computation occurs at the word level through word-level attention and feed-forward layers. Although byte-level attention and feed-forward layers substantially increase training and inference cost, EXECUTE’s Amharic results suggest they may be critical for strong character-level manipulation. Understanding their role is therefore key to optimizing the performance–efficiency trade-off.

In this work, we train models from scratch, enabling full control over model architecture and allowing us to answer these questions:

1. How do hierarchical models fare against pure byte-level models in character-level understanding?

2. Where is the optimal placement of the byte-level layers within the model?

3. Which is more critical for character-level understanding, character-level attention, or character-level feed-forward modules?

We start by comparing the hierarchical versus the pure byte-level, and explore additional, new strategies for creating a hierarchy. Our contributions are two-fold: 1) We find that a pure byte-level model is best for character-level understanding. 2) If the efficiency gains of a hierarchical model are desired, we present novel hierarchical strategies that better optimize the performance--efficiency trade-off. We provide code to reproduce our results.

2 Background

Hierarchical byte-level models aim to improve the efficiency of pure byte-level models while retaining byte-level input and output. Early approaches used convolutions and RNNs for downsampling and upsampling characters (Lee et al., 2017; Libovický et al., 2022; Edman et al., 2022, inter alia), while recent work uses transformers throughout (Slagle, 2024; Pagnoni et al., 2025; Minixhofer et al., 2026, inter alia). Most work focuses on how bytes are grouped into pseudo-word tokens, with strategies including constant, whitespace-, BPE-, and entropy-based sampling. Existing architectures generally follow the same pattern: downsample, process the shorter sequence, then upsample again, including multi-scale approaches such as Hourglass Transformers Nawrot et al. (2022).

Current state-of-the-art byte-level models include Byte Latent Transformer Pagnoni et al. (2025) and Bolmo Minixhofer et al. (2026). Both of these models have a similar architecture, using a single character-level transformer layer, followed by a large number (16-32) of word-level transformer layers, followed by a smaller number (4-9) for character-level decoder layers. These models perform better similarly to subword models on traditional benchmarks, while also better on character-level understanding tasks such as CUTE.

Character understanding is commonly evaluated with CUTE Edman et al. (2024), which tests simple character manipulations such as inserting random characters into words. Its multilingual extension, EXECUTE Edman et al. (2025), finds that models often perform better on low-resource scripts, likely because these scripts are processed more directly at the byte level and are less constrained by word-level biases.

Model	Notation
Byte	$(C_{a}C_{f})^{12}$
Top6	$(C_{a}C_{f})^{5}\,(W_{a}W_{f})^{6}\,(C_{a}C_{f})^{1}$
Mid6	$(C_{a}C_{f})^{3}\,(W_{a}W_{f})^{6}\,(C_{a}C_{f})^{3}$
Bot6	$(C_{a}C_{f})^{1}\,(W_{a}W_{f})^{6}\,(C_{a}C_{f})^{5}$
Intl6	$(C_{a}C_{f}\,W_{a}W_{f})^{3}\,(W_{a}W_{f}\,C_{a}C_{f})^{3}$
Mid8	$(C_{a}C_{f})^{2}\,(W_{a}W_{f})^{8}\,(C_{a}C_{f})^{2}$
Mid10	$(C_{a}C_{f})^{1}\,(W_{a}W_{f})^{10}\,(C_{a}C_{f})^{1}$
Mid10 CW	$(C_{a}C_{f})^{1}\,(C_{a}W_{f})^{10}\,(C_{a}C_{f})^{1}$
Mid10 WC	$(C_{a}C_{f})^{1}\,(W_{a}C_{f})^{10}\,(C_{a}C_{f})^{1}$

Table 1: 12-layer models tested and their notation. The number in the model name refers to the number of word-level layers.

	CUTE $\uparrow$			General $\uparrow$				Speed $\uparrow$		Mem $\downarrow$
	Pure	Diluted	Avg	ARC-E	Lambada	PIQA	Avg	Train	Infer	Train
Byte	94.0	31.2	62.6	42.0	19.8	61.7	41.2	1.00	1.00	1.00
Bot6	78.5	23.0	50.8	37.9	15.9	59.2	37.7	1.48	1.40	0.65
Top6	16.6	27.4	22.0	39.6	19.1	62.7	40.5	1.48	1.40	0.65
Intl6	10.2	15.2	12.7	40.3	19.1	62.1	40.5	1.48	1.40	0.65
Mid6	37.8	15.1	26.4	37.3	17.4	61.9	38.9	1.48	1.40	0.65
Mid8	27.1	14.0	20.5	39.8	18.8	62.8	40.5	1.76	1.63	0.54
Mid10	0.0	0.2	0.1	39.9	18.4	61.2	39.8	2.17	1.93	0.42
Mid10 CW	64.0	23.9	44.0	40.2	16.8	62.1	39.7	1.31	1.28	0.58
Mid10 WC	1.4	3.8	2.6	39.7	16.8	61.0	39.2	1.43	1.35	0.84
Subword	35.5	23.1	29.3	45.7	15.8	65.8	42.4	2.21	2.81	0.92

Table 2: Results on 12-layer models, on Pure and Diluted CUTE data, as well as standard benchmarks. We represent speed and memory as a ratio in relation to Byte.

3 Methodology

Given that our interest is in determining the architectural conditions for good character-level understanding, we develop a notation for the larger class of hierarchies we investigate.

3.1 Hierarchical Notation

Prior work on hierarchical byte-level models typically places word-level transformer layers either centrally, with similar numbers of character-level layers on both sides (Middle), or later in the model, with more character-level layers preceding them (Bottom). We also study a Top strategy, where most character-level layers come first, and an Interleaved strategy, where character- and word-level layers alternate (Figure 1).

Both attention and feed-forward modules can operate at either the byte or word level. Byte-level modules process all tokens, while word-level modules act only on masked subsets, with updates added back to the residual stream. Following Slagle (2024), we sample using whitespace delimiters, which works “surprisingly well in practice for English texts”.

We denote architectures using four module types: character-level attention ($C_{a}$), character-level feed-forward ($C_{f}$), word-level attention ($W_{a}$), and word-level feed-forward ($W_{f}$). Repeated patterns use exponents, e.g., $(C_{a}C_{f})^{12}$ for a standard 12-layer byte-level Transformer. Concatenation proceeds left-to-right, so Bolmo’s hierarchy is written as $(C_{a}C_{f})^{1}\,(W_{a}W_{f})^{N}\,(C_{a}C_{f})^{4}$, with $N=\{16,32\}$. Tables 1 and 4 list the 12- and 24-layer models tested, enabling control over word-layer placement, count, and module type. All models begin and end with fully character-level layers to ensure full token accessibility and avoid dead ends in the computation graph.

3.2 Experimental Setup

We evaluate two model scales: 12-layer models with hidden size 768 and 24-layer models with hidden size 1024, corresponding to roughly 100M and 400M parameters with a byte-level vocabulary. Following the scaling laws of Qiu et al. (2026), which suggest 9.6 GPT-2 tokens per parameter, we treat “GPT-2 tokens” as roughly equivalent to words and train on 1B- and 4B-word subsets of Nemotron-ClimbMix Diao et al. (2025) for the two scales.

The 1B-word experiments compare hierarchies and module types, while the larger-scale 4B experiments test whether the number of character-level layers should remain fixed or scale with model depth. Prior work differs on this choice: Slagle (2024) scaled character layers proportionally, whereas Minixhofer et al. (2026) kept them constant. We investigate which better supports character understanding.

We further continue training on a Diluted dataset combining additional ClimbMix data with 1% CUTE training data from Minixhofer et al. (2026), using 500 examples per subtask. This setup tests whether models can balance character-level and general language capabilities, similar to standard fine-tuning approaches.

We also train on a Pure dataset containing only CUTE data, again with 500 examples per subtask, to estimate the upper bound of character-understanding performance.

Training hyperparameters are reported in Appendix B. The longest run (a byte-level model on 4B words) required about five days on a single H100 GPU.

Evaluation focuses on CUTE, however we also report an average of standard NLP benchmarks: Lambada Paperno et al. (2016), PIQA Bisk et al. (2019), and ARC-Easy Clark et al. (2018). This is to assess whether architectural changes impact general NLP ability.

4 Results

We show our results on the 12-layer models in Table 2. We also include a subword model for reference, which uses the Olmo 2 vocabulary. In addition to CUTE and general benchmark scores, we show the training and inference speed, as well as training GPU memory usage, in relation to the Byte model. Overall, we see a significantly stronger performance for Byte on both the pure and diluted CUTE, while the general benchmarks have no clear winner, only mildly favoring Byte and Subword. We now break down the results according to our research questions.

Layer distribution

Comparing our different layering strategies (i.e., Bot, Mid, Top, Intl), the results indicate a preference for more character-level layers in the latter half of the transformer, with stronger performances from Bot and Mid on Pure CUTE. This tracks with the general understanding that these latter layers perform task-specific operations, where the ability to operate on the character-level would be more useful for CUTE. The diluted CUTE results appear more evenly balanced, suggesting the concentration of character-level layers on the top is only better when the task is focused.

Number of word-level layers

We increase the number of word-level layers from Mid₆ up to Mid₁₀. As the number of word-level layers increases, the performance drops. This is not surprising, as fewer layers tends towards the behavior of a subword model, but with the added difficulty of a small, byte-level vocabulary.

Attention vs. Feed-forward

The scores from Byte, Mid₁₀, Mid₁₀ CW, and Mid₁₀ WC paint a clear picture: attention is more important for character-level understanding. The two models with fully character-level attention, Byte and Mid₁₀ CW, substantially outperform those with mostly word-level attention.

	CUTE $\uparrow$			Speed $\uparrow$		Mem $\downarrow$
	Pure	Dilu.	Avg	Train	Infer	Train
Byte	97.5	55.8	76.6	1.00	1.00	1.00
Top12	91.7	51.2	71.5	1.51	1.45	0.65
Intl12	95.9	43.1	69.5	1.51	1.45	0.65
Intl18	91.8	44.3	68.1	2.04	1.87	0.47
Intl${}_{\text{hyb}}$	92.6	43.0	67.8	1.72	1.62	0.52
Mid12	94.3	40.7	67.5	1.51	1.45	0.65
Mid${}_{\text{hyb}}$	96.0	38.5	67.2	1.72	1.62	0.52
Bot${}_{\text{hyb}}$	93.0	40.4	66.7	1.72	1.62	0.52
Top18	91.4	38.7	65.0	2.04	1.87	0.47
Bot12	96.3	30.5	63.4	1.51	1.45	0.65
Top${}_{\text{hyb}}$	87.5	38.2	62.9	1.72	1.62	0.52
Mid18	91.6	26.4	59.0	2.04	1.87	0.47
Bot18	92.0	17.8	54.9	2.04	1.87	0.47
Subword	46.4	33.1	39.7	3.21	3.62	0.52

Table 3: Results on 24-layer models, sorted by average CUTE performance. We represent speed and memory as a ratio in relation to Byte.

Scaling Up

In Table 3, we present higher-resourced models ordered by average CUTE performance. Following the positive results of the $C_{a}W_{f}$ layer, we also evaluate hybrid architectures, including models with 6 $C_{a}W_{f}$ layers, 12 $W_{a}W_{f}$ layers, and 6 $C_{a}C_{f}$ layers (see Table 4 for details). We omit non-CUTE results for brevity; full results are in Appendix C.

Overall, the Byte model performs best by a 5% margin, showing stronger character understanding in both focused and general settings. We also note that the Byte model from the 1B setting, having roughly 8$\times$ less compute, performs better than Mid₁₈ and Bot₁₈ (62.6 vs. 59.0 and 54.9). In general, more character-level layers improve performance. On Pure, Mid and Bot perform strongly, but are lacking on Diluted, while Top₁₂ and Interleaved models are more robust in that setting. Hybrid models (apart from Top) show similar performance to those with 12 $C_{a}C_{f}$ layers, indicating $C_{a}W_{f}$ may be a cost-saving alternative to $C_{a}C_{f}$.

Despite training on only 4B words, nearly all models exceed 90% on Pure, substantially outperforming Bolmo despite its far larger training budget (49B adaptation tokens, 4T+ pretraining tokens, including 21M CUTE tokens). However, CUTE represents only 0.04% of its adaptation data, which explains its weaker retention of character-level knowledge.

Scaling trends show overall improvement, but architecture still matters. For example, Bot₁₈, which most closely resembles Bolmo and Byte Latent Transformer, improves only 4.1 points over Bot₆ when scaled 4$\times$ in model and data size. In contrast, Intl₁₈ improves 55.4 points, suggesting better scaling behavior and faster convergence toward the performance of Byte.

Overall, Byte models achieve the best performance, while Interleaved models offer a strong trade-off, enabling up to 2$\times$ faster training and reduced memory usage (less than half the RAM).

5 Conclusion

State-of-the-art byte-level models boast a better performance on character-level understanding tasks such as CUTE, however we have shown that achieving much better performance can be done without a hierarchical structure that downsamples to the word-level. We test several types of hierarchical structures, and all are outperformed by the fully byte-level models. However, due to their speed and memory efficiency, hierarchical models still may be desirable. In this case, we suggest an interleaved strategy, with alternating character-level and word-level layers.

There are still several open questions with byte-level models left unanswered. Concerning adaptation from a pretrained subword model (e.g., Bolmo from Olmo), an interleaved strategy may require a novel approach. Speeding up fully byte-level models would also be highly desirable.

Limitations

Our experiments are limited to English. We do not expect the results will differ in Latin-scripted languages, however languages such as Chinese may exhibit different trends due to the larger character set and shorter words. The whitespace sampling strategy we borrow from Slagle (2024) would not be suitable for some languages.

We also cannot exhaustively search all possible module configurations, so we limit the models tested to always alternating between attention and feedforward modules, as this is an established property of the Transformer architecture. However, even within this self-imposed limitation, we cannot test every possible configuration. We notably did not test models with a higher ratio of byte-level layers to word-level layers (e.g., Mid₂ or Mid₄). We expect these models to perform more and more similarly to the fully byte-level model, but they will consequently also have an increasingly similar speed and memory usage. As we have not seen any such models in prior research, we limit our study to models with an equal or higher word-to-character layer ratio.

Our experiments are on the 1B to 4B word scale, so we cannot definitively draw conclusions on very low resource or very high resource settings. Our trends indicate possible convergence of some of the hierarchical models to the byte-level performance, however it is not clear whether this indeed happens. We nevertheless show the capacity for better performance via our experiments on the Pure dataset.

Appendix A 24-layer models

We show the notation for 24-layer models in Table 4. The models follow a similar recipe to the 12-layer models (in Table 1), but adjusted to the 24-layer scale. Models denoted $X_{12}$ have 12 word-level layers, which is a 1:1 word/byte layer ratio. Models denoted $X_{18}$ have 18 word-level layers, with a 3:1 word/byte layer ratio, keeping the 6 byte layers constant from the 12-layer model experiments. Hybrid models are in between: 6 fully byte layers, 6 hybrid $C_{a}W_{f}$ layers, and 12 fully word layers.

Model	Notation
Byte	$(C_{a}C_{f})^{24}$
Top12	$(C_{a}C_{f})^{11}\,(W_{a}W_{f})^{12}\,(C_{a}C_{f})^{1}$
Top18	$(C_{a}C_{f})^{5}\,(W_{a}W_{f})^{18}\,(C_{a}C_{f})^{1}$
Top${}_{\text{hyb}}$	$(C_{a}C_{f})^{5}\,(C_{a}W_{f})^{6}\,(W_{a}W_{f})^{12}\,(C_{a}C_{f})^{1}$
Mid12	$(C_{a}C_{f})^{6}\,(W_{a}W_{f})^{12}\,(C_{a}C_{f})^{6}$
Mid18	$(C_{a}C_{f})^{3}\,(W_{a}W_{f})^{18}\,(C_{a}C_{f})^{3}$
Mid${}_{\text{hyb}}$	$(C_{a}C_{f})^{3}\,(C_{a}W_{f})^{3}\,(W_{a}W_{f})^{12}\,(C_{a}W_{f})^{3}\,(C_{a}C_{f})^{3}$
Bot12	$(C_{a}C_{f})^{1}\,(W_{a}W_{f})^{12}\,(C_{a}C_{f})^{11}$
Bot18	$(C_{a}C_{f})^{1}\,(W_{a}W_{f})^{18}\,(C_{a}C_{f})^{5}$
Bot${}_{\text{hyb}}$	$(C_{a}C_{f})^{1}\,(W_{a}W_{f})^{12}\,(C_{a}W_{f})^{6}\,(C_{a}C_{f})^{5}$
Intl12	$(C_{a}C_{f}\,W_{a}W_{f})^{6}\,(W_{a}W_{f}\,C_{a}C_{f})^{6}$
Intl18	$(C_{a}C_{f}\,(W_{a}W_{f})^{3})^{2}\,((W_{a}W_{f})^{3}\,C_{a}C_{f})^{2}$
Intl${}_{\text{hyb}}$	$(C_{a}C_{f}\,W_{a}W_{f}\,C_{a}W_{f}\,W_{a}W_{f})^{2}\,(W_{a}W_{f}\,C_{a}W_{f}\,W_{a}W_{f}\,C_{a}C_{f})^{2}$

Table 4: 24-layer models tested and their notation. The number in the model name refers to the number of word-level layers.

Appendix B Hyperparameters

We show the hyperparameters used in our experiments in Table 5. We vary the batch size in our ClimbMix experiments to keep the number of iterations roughly equal. In initial testing, we used a learning rate of $4e-4$, following Byte Latent Transformer, but scaled down to $2e-4$ due to the smaller scales requiring smaller batch sizes.

Dataset	ClimbMix 1B	ClimbMix 4B	Diluted	Pure
Batch size	32	128	32	8
Sequence length	2048	2048	2048	2048
Epochs	1	1	5	5
LR	2e-4	2e-4	1e-4	2e-5
Warmup steps	2000	2000	2000	50
Optimizer	AdamW	AdamW	AdamW	AdamW
Betas	(.9, .98)	(.9, .98)	(.9, .98)	(.9, .999)
Weight decay	0.01	0.01	0.01	0.01
Grad norm clip	1	1	1	1
Precision	bf16	bf16	bf16	bf16

Table 5: Hyperparameters used for our experiments.

Appendix C Complete Results

All of the results obtained in our experiments are presented in Table 6. Concerning the CUTE subtasks, we also split them based on character- versus word-level tasks, but found no meaningful difference in the trends. We suspect this is because all of the models tested are byte-level, so the word-level tasks of CUTE can be thought of as several sequential byte-level operations.

Size	Dataset	Model	Spell	Spell Inv	Cont C	Cont W	Ins C	Ins W	Del C	Del W	Sub C	Sub W	Swap C	Swap W	ARC-E	Lambada	PIQA
1B	Diluted	Byte	70.9	53.6	64.8	64.6	12.4	7.2	28.0	24.5	21.8	17.7	6.4	2.1	42.0	19.8	61.7
1B	Diluted	Bot6	70.3	50.6	57.9	52.4	2.1	3.7	6.7	6.8	10.5	6.3	6.3	2.6	37.9	15.9	59.2
1B	Diluted	Top6	36.7	53.3	51.2	57.5	6.0	13.0	33.8	16.2	36.8	16.5	5.1	2.4	39.6	19.1	62.7
1B	Diluted	Intl6	20.4	10.5	38.9	51.3	0.9	8.8	10.4	16.2	10.4	13.1	0.5	0.9	37.3	17.4	61.9
1B	Diluted	Mid6	5.6	17.4	58.0	62.1	3.2	1.4	6.4	2.3	14.9	4.0	5.3	0.4	40.3	19.1	62.1
1B	Diluted	Mid8	0.2	33.2	49.3	53.1	0.9	1.4	3.1	5.3	15.3	3.9	2.4	0.0	39.8	18.8	62.8
1B	Diluted	Mid${}_{1}0$	0.0	0.3	0.0	0.0	0.0	0.0	0.0	1.1	0.7	0.3	0.0	0.0	39.9	18.4	61.2
1B	Diluted	Mid10 CW	54.2	67.7	56.2	65.5	1.8	7.5	7.3	8.2	9.4	8.5	0.4	0.5	40.2	16.8	62.1
1B	Diluted	Mid10 WC	0.0	9.6	5.2	21.5	0.1	0.4	0.9	4.6	2.0	1.1	0.1	0.0	39.7	16.8	61.0
1B	Pure	Byte	99.2	99.5	100.0	99.4	91.4	90.8	97.7	90.2	98.5	93.5	95.4	71.9	-	-	-
1B	Pure	Bot6	99.6	99.1	69.9	81.8	61.4	80.1	91.1	78.9	92.4	71.1	81.9	35.3	-	-	-
1B	Pure	Top6	0.1	40.9	24.4	19.3	8.1	0.1	39.2	0.1	55.9	0.0	10.6	0.0	-	-	-
1B	Pure	Intl6	40.7	88.6	5.2	1.5	70.2	0.0	88.4	4.7	85.1	1.0	67.8	0.0	-	-	-
1B	Pure	Mid6	0.0	31.7	0.0	1.6	14.6	0.0	21.8	1.2	46.1	0.1	4.8	0.0	-	-	-
1B	Pure	Mid8	8.6	58.4	0.0	0.0	76.1	0.0	68.7	3.0	71.5	1.9	36.7	0.0	-	-	-
1B	Pure	Mid${}_{1}0$	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	-	-	-
1B	Pure	Mid10 CW	98.6	99.0	69.1	79.5	29.5	71.5	76.2	65.1	76.7	53.7	31.7	17.6	-	-	-
1B	Pure	Mid10 WC	0.0	7.1	0.0	0.0	2.3	0.1	0.9	0.0	6.9	0.0	0.1	0.0	-	-	-
4B	Diluted	Byte	73.3	86.4	65.6	69.6	33.6	58.3	59.1	53.4	60.8	62.1	25.3	21.7	46.4	25.4	65.6
4B	Diluted	Bot12	33.0	11.1	54.1	66.3	9.9	29.5	46.3	27.0	21.9	21.9	29.2	15.3	45.7	23.0	65.1
4B	Diluted	Bot18	10.8	14.1	46.3	54.4	0.2	6.4	47.9	12.5	6.7	4.6	8.3	2.0	48.6	23.3	66.1
4B	Diluted	Bot${}_{\text{hyb}}$	30.4	63.5	54.0	75.4	7.2	19.1	62.6	29.2	51.9	25.0	54.2	12.3	48.6	21.7	65.7
4B	Diluted	Mid12	77.0	75.4	20.4	69.5	20.0	31.3	46.7	24.3	57.6	35.0	22.2	8.5	48.2	25.3	65.1
4B	Diluted	Mid18	19.0	53.3	59.3	74.9	10.6	7.6	18.4	6.3	31.4	26.7	8.3	1.3	49.6	22.3	66.8
4B	Diluted	Mid${}_{\text{hyb}}$	66.6	79.5	63.3	63.9	12.3	13.5	29.8	28.9	43.5	41.6	9.1	9.5	48.8	23.7	66.8
4B	Diluted	Intl12	46.9	86.8	65.4	78.6	22.8	55.3	55.2	44.0	63.6	55.9	19.2	20.6	45.8	20.9	65.6
4B	Diluted	Intl18	54.9	77.5	60.8	59.7	12.3	44.2	28.0	31.0	28.5	46.0	2.1	0.0	48.1	23.0	64.0
4B	Diluted	Intl${}_{\text{hyb}}$	4.5	66.1	65.6	59.8	9.0	48.4	39.0	38.4	46.0	50.8	16.5	14.7	47.1	23.2	66.6
4B	Diluted	Top12	63.3	61.5	67.9	66.6	22.3	45.6	42.1	28.0	41.7	49.9	16.2	12.3	46.3	23.4	64.4
4B	Diluted	Top18	78.4	80.9	57.7	65.4	9.8	37.7	40.2	26.3	61.6	40.6	20.9	12.3	46.1	22.5	65.4
4B	Diluted	Top${}_{\text{hyb}}$	77.3	61.7	64.2	71.0	7.9	38.5	41.5	13.0	72.1	20.1	42.3	0.0	48.4	21.7	64.5
4B	Pure	Byte	99.9	99.7	99.8	99.7	95.0	96.6	99.4	94.4	98.6	97.4	97.1	92.8	-	-	-
4B	Pure	Bot12	100.0	99.9	99.9	99.9	96.2	95.5	98.2	90.8	99.2	94.5	98.4	82.5	-	-	-
4B	Pure	Bot18	100.0	99.5	72.2	99.7	90.4	96.6	95.7	90.3	94.9	95.4	84.4	84.9	-	-	-
4B	Pure	Bot${}_{\text{hyb}}$	99.8	100.0	70.3	99.5	90.4	95.7	95.2	90.8	98.5	95.0	96.6	84.8	-	-	-
4B	Pure	Mid12	99.9	99.7	99.7	100.0	92.4	95.6	96.8	91.1	94.8	96.9	77.2	87.3	-	-	-
4B	Pure	Mid18	99.2	99.7	71.3	99.9	92.3	95.7	94.2	89.6	92.7	96.7	79.7	87.8	-	-	-
4B	Pure	Mid${}_{\text{hyb}}$	99.1	99.6	99.4	99.4	93.6	95.8	97.7	92.0	98.3	97.4	93.1	86.5	-	-	-
4B	Pure	Intl12	99.8	99.6	99.4	99.5	93.9	95.2	98.2	91.4	97.1	97.8	91.0	88.3	-	-	-
4B	Pure	Intl18	99.2	98.8	99.7	99.7	87.4	94.5	94.6	89.9	91.7	95.6	67.5	83.1	-	-	-
4B	Pure	Intl${}_{\text{hyb}}$	99.5	99.1	77.8	99.2	89.4	93.6	91.5	90.2	97.7	95.8	95.1	81.9	-	-	-
4B	Pure	Top12	99.6	99.4	98.9	99.9	86.9	91.5	92.9	86.4	92.5	94.8	79.0	78.8	-	-	-
4B	Pure	Top18	98.6	99.2	99.8	99.9	78.8	92.3	90.1	88.6	93.2	95.5	77.4	83.1	-	-	-
4B	Pure	Top${}_{\text{hyb}}$	99.2	99.7	69.6	99.8	71.3	91.6	85.1	87.9	91.5	93.3	80.0	80.7	-	-	-

Table 6: Complete results.