The Dutch East India Company and the VOC archives

Established in 1602, the VOC dominated early modern maritime commerce and colonial administration, particularly in Asia (Stoler Reference Stoler2002). Its surviving records, millions of pages of correspondence, ledgers and court proceedings, constitute a uniquely detailed source of economic, social and political history (Petram and Van Rossum Reference Petram and Van Rossum2022). Although large-scale digitization has facilitated access (Manjavacas and Fonteyn Reference Manjavacas and Fonteyn2022) to these records, computational use of the corpus remains limited because the material is written in early modern Dutch. The variety diverges markedly from the present Dutch in vocabulary, syntax and idiom (Jiang and Zhang Reference Jiang and Zhang2024), and there is no large bilingual corpus to train conventional neural-machine translation (NMT) systems (Huang et al. Reference Huang, Zhang, Geng, Yichao, Chen, Huang, Lun-Wei, Martins and Srikumar2024). Specialized fine-tuning of multilingual LLMs therefore offers a promising, but still underexplored, pathway to research accessibility.

Testimonies of Enslavement

“Testimonies of Enslavement” comprises English translations of court cases heard by the VOC’s Court of Justice in Cochin (Kerala) during 1680–1682 and 1707–1792 (van Rossum et al. Reference van Rossum, Geelen, Van den Hout and Tosun2020). This unique collection presents parallel texts in Dutch, Portuguese, Latin and local languages, systematically expanding historical abbreviations while preserving culturally specific legal and administrative terms that resist literal translation, precisely the linguistic challenges that complicate automated translation of Early Modern Dutch. The corpus’s digitization includes detailed metadata covering case type, date, thematic tags and personal attributes, such as gender or legal status, making it a benchmark for studying slavery, migration and cross-cultural jurisprudence. Most critically for machine translation (MT) research, the collection’s sentence-aligned parallel structure provides the expert-curated bilingual data essential for fine-tuning and evaluating historical translation models, bridging the gap between raw archival material and the structured datasets required for NMT.

Fine-tuning LLMs for specialized translation

MT remains a flagship application of transformer-based LLMs (Vaswani et al. Reference Vaswani, Shazeer, Parmar, Guyon, von Luxburg, Bengio, Wallach, Fergus, Vishwanathan and Garnett2017). NMT systems outperform earlier rule-based and statistical approaches because they learn cross-lingual patterns from large parallel corpora (Huang et al. Reference Huang, Zhang, Geng, Yichao, Chen, Huang, Lun-Wei, Martins and Srikumar2024). A full fine-tune updates every weight in a multi-billion-parameter network, which demands expensive hardware. PEFT instead adjusts only a tiny slice of the model (often <1%), so researchers can adapt an LLM on a laptop-grade GPU while leaving the core weights untouched. Fine-tuning adapts these general models to low-resource historical domains through two successive steps:

1. 1. Instruction tuning – also called SFT, exposes the model to curated sentence pairs so that it acquires the orthography, vocabulary and syntax of early modern Dutch.

2. 2. Preference alignment reshapes the model’s output style using human feedback. Popular methods include reinforcement learning from human feedback (RLHF) (Ouyang et al. Reference Ouyang, Jeff, Jiang, Alice, Naumann, Globerson, Saenko, Hardt and Levine2022), direct preference optimization (DPO) (Rafailov et al. Reference Rafailov, Sharma, Mitchell, Ermon, Manning, Finn, Alice, Naumann, Globerson, Saenko, Hardt and Levine2023), lightweight LoRA adapters (LoRA adds a small trainable matrix to each layer, enabling parameter-efficient updates.) (Hu et al. Reference Hu, Shen, Wallis, Allen-Zhu, Li, Wang, Lu and Chen2022) and ORPO (Hong, Lee, and Thorne Reference Hong, Lee, Thorne, Salakhutdinov, Kolter and Heller2024).

Combined, these stages yield a translator that preserves historical nuance while meeting present-day quality expectations.

Machine translation

MT enables automatic transfer of text across languages and is indispensable for unlocking historical sources (Sennrich, Haddow, and Birch Reference Sennrich, Haddow, Birch, Erk and Smith2016). Recent advances in NMT have markedly improved adequacy and fluency, even for linguistically complex and culturally embedded material such as early modern Dutch (Gao et al. Reference Gao, Ma, Lin, Callan, Rogers, Boyd-Graber and Okazaki2022).

Three methodological stages frame the field:

1. 1. Rule-based MT: Hand-crafted transfer rules and bilingual lexicons performed reliably in restricted domains, but proved brittle with orthographic variation, archaic morphology and long-distance dependencies (Wang et al. Reference Wang, Hua, He, Huang and Church2022).

2. 2. Statistical MT: Word and phrase-based probabilistic models (Brown, Della, and Pietra Reference Brown, Della and Pietra1993; Lopez Reference Lopez2008) replaced rules with corpus statistics, improving robustness but still struggling with reordering and rich morphology: complex inflectional patterns where words change form for grammatical functions, creating numerous variants that fragment translation probabilities.

3. 3. NMT: Sequence-to-sequence architectures with attention (Bahdanau, Cho, and Bengio Reference Bahdanau, Cho, Bengio, Bengio and LeCun2015), culminating in the Transformer (Vaswani et al. Reference Vaswani, Shazeer, Parmar, Guyon, von Luxburg, Bengio, Wallach, Fergus, Vishwanathan and Garnett2017), delivered large leaps in quality. Sub-word segmentation (Sennrich, Haddow, and Birch Reference Sennrich, Haddow, Birch, Erk and Smith2016) and massively multilingual pre-training (Aharoni, Johnson, and Firat Reference Aharoni, Johnson, Firat, Burstein, Doran and Solorio2019; Zhang and Sennrich Reference Zhang, Sennrich, Jurafsky, Chai, Schluter and Tetreault2020) further widened language coverage.

Unlike traditional MT systems that are explicitly trained on parallel corpora for translation tasks, contemporary LLMs acquire translation capabilities emergently through their general language modeling objectives. While specialized NMT systems remain optimized for translation quality, LLMs offer a more flexible approach that can handle multiple tasks, including translation, within a single model architecture. This study bridges these paradigms by adapting general-purpose LLMs to specialized historical translation through PEFT, combining the flexibility of LLMs with the domain specificity of traditional MT approaches.

• • Full fine-tuning updates every parameter, achieving state-of-the-art accuracy but incurring high memory cost and risking catastrophic forgetting.

• • PEFT freezes the backbone and trains lightweight modules – adapters (Houlsby et al. Reference Houlsby, Giurgiu, Jastrzebski, Morrone, De Laroussilhe, Gesmundo, Attariyan, Gelly, Chaudhuri and Salakhutdinov2019), prefix-tuning (Li and Liang Reference Li, Liang, Zong, Xia, Li and Navigli2021) or LoRA matrices (Hu et al. Reference Hu, Shen, Wallis, Allen-Zhu, Li, Wang, Lu and Chen2022) that reduce compute without sacrificing quality.

LLMs offer two further advantages for historical MT: (i) zero- or few-shot performance on extremely low-resource pairs and (ii) unified multitask capability translation, summarization and extraction in one model. This study therefore tests two PEFT strategies, Unsloth SFT and ORPO, to inject early modern Dutch expertise into open-source LLMs while keeping training and inference budgets modest. The resulting models are benchmarked with embedding-based metrics that reward semantic fidelity and stylistic precision.

Translation evaluation

Given the complexity and historical significance of translating early modern Dutch texts, accurate evaluation is crucial. This study scores system outputs with two automatic metrics chosen for their ability to capture semantic adequacy and fluency:

(i) BERTScore (Zhang et al. Reference Zhang, Kishore and Felix2020), a semantic-similarity metric that leverages contextual embeddings to gauge how closely a hypothesis preserves the reference meaning, an essential property for historically nuanced material.

(ii) METEOR (Banerjee and Lavie Reference Banerjee, Lavie, Goldstein, Lavie, Lin and Voss2005; Denkowski and Lavie Reference Denkowski, Lavie, Bojar, Buck, Federmann, Haddow, Koehn, Leveling, Monz, Pecina, Post, Saint-Amand, Soricut, Specia and Tamchyna2014), which aligns tokens at the exact, stem, synonym and paraphrase levels to balance lexical richness, fluency and stylistic authenticity.

Together these metrics overcome tokenization mismatches and length bias that affect traditional n-gram measures such as BLEU (Papineni et al. Reference Papineni, Roukos, Ward, Zhu and Isabelle2002; Post Reference Post2018), aligning with the goal of robust, historically-aware MT evaluation.

BERTScore

BERTScore compares a candidate sentence $C=\{c_1,\dots ,c_{|C|}\}$ with a reference sentence $R=\{r_1,\dots ,r_{|R|}\}$ in a contextual-embedding space. Each token is encoded by the multilingual bert-base-multilingual-cased model (Devlin et al. Reference Devlin, Chang, Lee, Toutanova, Burstein, Doran and Solorio2019) (layer 8 by default). Cosine similarities form the matrix $S_{ij}=\cos \!\bigl (\mathbf {e}(c_i),\mathbf {e}(r_j)\bigr )$ . Keeping the maximum similarity for each token yields

$$\begin{align*}\text{Precision}= \frac{1}{|C|}\sum_{i}\max_j S_{ij}, \qquad \text{Recall}= \frac{1}{|R|}\sum_{j}\max_i S_{ij}. \end{align*}$$

The harmonic mean gives the final score,

$$\begin{align*}\text{BERTScore}= 2\, \frac{\text{Precision}\times\text{Recall}} {\text{Precision}+\text{Recall}}, \end{align*}$$

after inverse-document-frequency weighting that down-weights very common words. Sentence-level scores are averaged to produce the system-level values reported in the “Results” section (range 0–1, where 1 denotes identical meaning).

METEOR

METEOR aligns hypothesis and reference tokens via exact, stem, synonym and paraphrase matches (Banerjee and Lavie Reference Banerjee, Lavie, Goldstein, Lavie, Lin and Voss2005). Let m be the number of matched tokens; precision (P) and recall (R) are

$$\begin{align*}P=\frac{m}{|C|}, \qquad R=\frac{m}{|R|}, \quad F_{\!mean}= \frac{10\,P\,R}{R+9P}. \end{align*}$$

Fragmentation incurs a penalty,

$$\begin{align*}\text{Penalty}=0.5 \bigl(\tfrac{\text{chunks}}{m}\bigr)^{3}, \qquad \text{METEOR}=F_{\!mean}\,(1-\text{Penalty}), \end{align*}$$

yielding a score between 0 and 1. We use METEOR v1.5 (SacreBLEU implementation) with the English WordNet synonym list and the paraphrase table of Denkowski and Lavie (Reference Denkowski, Lavie, Bojar, Buck, Federmann, Haddow, Koehn, Leveling, Monz, Pecina, Post, Saint-Amand, Soricut, Specia and Tamchyna2014).

Tokenization

Tokenization is the initial parsing step in most NLP pipelines: It segments raw text into words, subwords or bytes that LLMs can process (Mielke et al. Reference Mielke, Alyafeai and Salesky2021). For multilingual or historical LLMs, the chosen scheme directly influences how well the network captures vocabulary, morphology and syntax (Kudo and Richardson Reference Kudo, Richardson, Blanco and Lu2018; Manjavacas and Fonteyn Reference Manjavacas and Fonteyn2022; Sennrich, Haddow, and Birch Reference Sennrich, Haddow, Birch, Erk and Smith2016); translation quality therefore hinges on tokenization fidelity (Hu et al. Reference Hu, Jain, Elmoznino, Kaddar, Lajoie, Bengio and Malkin2024).

Efficient tokenizers compress more information into each context window, reducing training time and inference cost (Jiang and Zhang Reference Jiang and Zhang2024). Early modern Dutch, however, exhibits variable spelling, archaic graphemes and mixed scripts that tokenizers optimized for present-day corpora often fragment. The resulting high out-of-vocabulary (OOV) rates lengthen sequences, split morpho-syntactic units and depress model performance (Manjavacas and Fonteyn Reference Manjavacas and Fonteyn2022; Mielke et al. Reference Mielke, Alyafeai and Salesky2021). Targeted vocabulary augmentation and bespoke subword training are, therefore, prerequisites for accurate translation and analysis of historically significant documents. In this study, a custom tokenizer was developed to show what optimal processing of our source corpus would look like, the performance of which is reported as part of the methodological setup.

Tokenizer implementation

To mitigate the issues of orthographic fragmentation outlined in the “Tokenization” section, a custom byte-pair encoding (BPE) tokenizer was trained on the Early Modern Dutch corpus. This step served as a foundational preprocessing measure to ensure the subsequent fine-tuning models could efficiently process the historical texts. Its performance was compared against several standard tokenizers (GPT-2, mBERT, and XLM-R) on the test set. As shown in Table 1, the custom tokenizer achieved an impressive OOV coverage in contrast to other tokenizers and the highest sequence efficiency, representing the source texts in the fewest number of tokens. This optimization reduced computational overhead and potential context fragmentation during the fine-tuning and inference stages.

Table 1.

Tokenizer performance comparison on early modern Dutch

Testing environment

All inference ran in Google Colab on NVIDIA L4 and A100 GPUs via a custom notebook. (Notebook archive: https://anony mous.4open.science/r/EMDutchLLMFinetuneA67B.) Each model received a one-shot prompt consisting of (i) an instruction to render an early modern Dutch sentence in contemporary English under specified guidelines and (ii) the sentence itself (Lip Reference Lip2025). Generation parameters were fixed across systems: do_sample=True, num_beams=5, penalty_alpha=0.6. The models translated 230 sentences from 18th-century court cases; outputs were written to CSV for subsequent scoring.

Quantitative benchmark

Python implementations of BERTScore and METEOR (Lip Reference Lip2025) produced scores for every hypothesis–reference pair. BERTScore captures sentence-level semantic similarity, while METEOR assesses lexical choice, phrase alignment and fluency. Their conjunction provides a more complete picture of translation quality than either metric alone.

Qualitative benchmark

For the qualitative benchmark, a domain expert performed a systematic review of all 230 test sentences across all model outputs. Rather than aiming for “perfect” automated translations – an unrealistic goal given the complexity of historical texts – our framework positions model outputs as preliminary drafts for expert refinement. In practice, historians would work with these imperfect translations, using them as starting points that significantly accelerate the traditional translation process while maintaining scholarly oversight.

Base models

Experiments start from publicly available, instruction-tuned checkpoints:

• • Llama-3-8B (Meta) – SFT and ORPO;

• • Llama-2-7B (Meta) – SFT;

• • Mistral-7B-Instruct-v0.3 (MistralAI) – SFT;

• • TinyLlama-1.1B – SFT.

Compute environment

Fine-tuning was performed in Google Colab Pro on NVIDIA L4 (24 GB) and A100 (40 GB) GPUs. Notebooks and configuration files are archived in the project repository (Lip Reference Lip2025). Training curves were logged to Weights and Biases for real-time monitoring (Biewald Reference Biewald2020).

Hyper-parameters

All runs use an accumulative batch size of 64 and a maximum sequence length of 512 tokens. For the full parameter list for each model, see Table A6.

• • SFT (Unsloth): Nine epochs, AdamW optimizer, learning rate $2\times 10^{-4}$ , $\beta _{1}=0.9$ , $\beta _{2}=0.999$ .

• • ORPO: Eight epochs, AdamW optimizer, learning rate $1\times 10^{-5}$ , same $\beta $ values, KL-divergence weight $0.02$ .

Fine-tuned model capabilities

The fine-tuned systems adapted with the Unsloth procedure markedly outperformed their baseline counterparts. Mistral-Unsloth achieved the closest alignment with expert reference translations, confirming that Unsloth can capture the lexical nuance and syntactic irregularity characteristic of early modern Dutch. Its low score variance further recommends it for applications in which scholars require dependable, minimally post-edited output.

Llama-3-Unsloth also profited from the same adaptation, although its wider dispersion of scores indicates residual sensitivity to sentence complexity and specialized terminology. Even so, the model’s mean performance approaches that of Mistral-Unsloth, shown in Figure 5, underscoring the practical value of targeted fine-tuning for historical translation. Together, the two models demonstrate that parameter-efficient strategies can yield both accuracy and stability, provided the backbone has sufficient capacity and the domain data are carefully curated.

Figure 5.

Sentence-level METEOR distributions for the two best systems.

ORPO fine-tuning challenges

The Llama-3-8B model fine-tuned with ORPO delivered markedly lower and less reliable scores than the Unsloth SFT baselines. Its mean BERTScore was 0.617 (median = 0.665, SD = 0.180) and its mean METEOR 0.405 (median = 0.461, SD = 0.246; see Tables A10 and A11). Although the model occasionally produced near-expert translations, the wide dispersion of scores indicates that such high-quality outputs were unpredictable and infrequent. A review of the training notebook and data flow points to three primary implementation issues that likely caused this underperformance, rather than an intrinsic flaw in the ORPO algorithm for this task.

1. a) Sparse, imbalanced preference pairs: ORPO’s efficacy depends on a sufficient volume of high-quality chosen-rejected example pairs. With only 1,281 sentence-translation tuples available, creating the preference dataset effectively halved the number of examples seen per epoch compared to SFT. This data scarcity forces the model to learn from a very narrow distribution, increasing the risk of overfitting to frequent lexical patterns in the training data rather than learning generalizable syntactic and semantic rules for translation. Furthermore, using a powerful model like GPT-4o for the “rejected” samples, while a common practice, may have introduced high-quality but stylistically different translations, potentially sending a noisy or conflicting signal to the model during alignment.

2. b) Prompt-template drift: A critical mismatch existed between the structured, multi-line instruction block used during training and the simple, single-sentence prompt used at inference. Instruction-tuned models are highly sensitive to the format of the prompt. This discrepancy meant that the model was not optimized for the task it was being asked to perform during evaluation. Consequently, the model often exhibited prompt leakage, reproducing parts of the training header or failing to complete the translation (shown in Table A5), which severely depressed the adequacy scores and contributed to the high variance in performance.

3. c) Conservative hyper-parameters: The training configuration, which ran for eight epochs with a low learning rate of $1\times 10^{-5}$ and a micro-batch size of two, was likely too conservative for this context. In ORPO, the language modeling loss is balanced by a KL-divergence regularizer that prevents the model from deviating too far from its original capabilities. When combined with a low learning rate, this regularizer can dominate the training process, effectively stifling the model’s ability to adapt to the new domain. The model is thus prevented from fully learning the nuances of Early Modern Dutch, leading to translations that may remain closer to the base model’s generic output rather than acquiring specialized historical knowledge.

These three factors – data sparsity, prompt inconsistency and a restrictive training regime – collectively explain the 0.20–0.35 drop in mean scores and the nearly three-fold increase in variance relative to the Unsloth SFT models, without necessarily indicting ORPO’s underlying algorithm for historical translation tasks.

Base model capabilities

The untuned Mistral-7B-Instruct-v0.3 delivered stable BERTScore and METEOR values across the test set, confirming that it possesses a strong multilingual prior even without domain adaptation. This consistency makes it an attractive foundation for historical translation workflows. Once fine-tuned with Unsloth SFT the model’s average scores rose by 8–10 points, and variance fell by roughly one third, indicating both higher accuracy and greater reliability (Table A10). The gain demonstrates the practical value of lightweight parameter-efficient tuning when adapting a general LLM to the lexical and syntactic idiosyncrasies of 17th- and 18th-century Dutch.

Among the other baselines, GPT-4o produced the most dependable high-quality output without any additional training. Although its mean scores lagged slightly behind the Unsloth-enhanced Mistral, the narrow dispersion of GPT-4o shows that it can serve as a robust standard alternative when compute or data for fine-tuning are unavailable (Figures 3 and 4). Taken together, these results suggest a two-tier deployment strategy: use GPT-4o as an immediate, low-overhead baseline and adopt a fine-tuned Mistral backend when project resources permit domain adaptation and post-editing.

Discussion

The comparative results underline how strongly translation quality varied between fine-tuning strategies. This is shown starkly in Figure A1, which highlights the jump in performance from base model to fine-tuned model through a heatmap visualization. Unsloth adaptation, exemplified by the Mistral-Unsloth and Llama-3-Unsloth models (Figures 3 and 4), consistently surpassed all alternatives. The ORPO variant remains less successful, but its shortcomings appear remediable (“ORPO fine-tuning challenges” section). Adjusting the preference corpus and harmonizing training and inference prompts may narrow the current performance gap. More broadly, combining automatic metrics with an authenticity checklist covering proper-noun preservation, date fidelity and punctuation style would yield a balanced scorecard for future experiments. A recurring error pattern across systems is the handling of very short lexical items and abbreviations in Dutch: when a sentence is both brief and semantically opaque, the model often leaves such tokens untranslated because their functional role is unclear. Targeted glossaries or context-window augmentation could mitigate this issue without sacrificing fluency.

Historical translators must decide which elements should remain unchanged to preserve documentary authenticity. Proper names, official titles (schout, fiscaal) and obsolete currency units (e.g., stuiver) are prime examples. Our expert evaluation shows that Mistral-Unsloth retains almost all personal and place names verbatim, whereas Llama-3-Unsloth occasionally anglicizes Dutch surnames or expands abbreviated titles. In contrast, GPT-4o sometimes over-translates institutional names, rendering Hoogheemraadschap as “Water Board,” which may help modern readers but risks erasing historical nuance. A pragmatic compromise is to keep these critical tokens in their original form and provide a parenthetical gloss on the first occurrence. Implementing this policy would require a lightweight named-entity filter upstream of the decoder or a constrained decoding pass that protects whitelisted terms.

Findings

This study compared two parameter-efficient adaptation strategies, ORPO and Unsloth SFT for translating early modern Dutch into present-day English assessed through both qualitative and quantize frameworks. Both methods improved substantially on untuned foundation models, yet their performance profiles diverge in ways that matter for historical scholarship. Unsloth fine-tuned Mistral and Llama 3 systems delivered the highest BERTScore and METEOR values while preserving most syntactic and stylistic features observed in expert references. ORPO, by contrast, produced more variable output: when reward signals aligned with domain needs, it matched Unsloth quality, but miscalibrated preferences yielded prompt leakage or unfinished clauses. The TinyLlama experiment shows that extreme parameter reduction comes at a steep cost to adequacy and fluency, confirming that complex, low-resource language varieties still demand mid-size or larger backbones. Taken together, the results advocate for a two-step workflow, tokenizer customization, followed by Unsloth SFT for projects that require dependable, scalable translation of historical texts.

Scalable infrastructure

The present experiments were limited by GPU memory, restricting model depth, batch size and context window. Even a modest increase in compute would enable larger backbones and longer sequences; scaling-law studies indicate near-linear gains in accuracy under such conditions (Hoffmann et al. Reference Hoffmann, Borgeaud, Mensch, Alice, Naumann, Globerson, Saenko, Hardt and Levine2022; Kaplan et al. Reference Kaplan, McCandlish and Henighan2020). Beyond simply improving accuracy, scaling up the infrastructure would allow for fundamentally new research avenues. For instance, larger context windows (e.g., 32k or 128k tokens) would enable the translation of entire documents or chapters at once, rather than sentence by sentence. This would significantly improve coherence and consistency, allowing the model to track entities, terminology and narrative threads across long passages. Furthermore, employing larger backbone models (e.g., 70B or mixture-of-experts models) could enhance the system’s ability to grasp complex legal reasoning, subtle irony or intricate rhetorical structures present in the source texts, which smaller models might oversimplify.

Richer, well-curated data

Translation quality is ultimately capped by the scarcity of parallel material. Expanding the corpus across genres and annotating each record with place, year and document type remain a priority. Automatic workflows that recognize 17th- and 18th-century typographic conventions can accelerate transcription, while a tokenizer retrained on the enlarged dataset should capture historical spelling variation without inflating sequence length (Kudo and Richardson Reference Kudo, Richardson, Blanco and Lu2018). While the current corpus of court testimonies is valuable, its legalistic and formal nature may bias the model. Future work should prioritize expanding the training data to include a wider array of genres, such as personal letters, ship logs, administrative ledgers and pamphlets. This would create a more versatile translator capable of handling different registers and domains. Additionally, a significant challenge is the lack of a standardized transcription protocol for Early Modern Dutch. A key future task will be to develop and apply a semi-automated transcription workflow that can handle abbreviations, non-standard spelling and paleographic variations, perhaps using a combination of OCR/HTR models fine-tuned on specific scribal hands. This would not only accelerate corpus creation but also enrich the dataset with valuable metadata about scribal practices. This expanded and diversified corpus would, in turn, serve as the foundation for retraining the domain-specific tokenizer, further enhancing its ability to model historical spelling variation across genres and time periods.

Enhanced alignment strategies

ORPO lagged behind SFT chiefly because of sparse preference pairs and divergent prompt templates. Forthcoming work will (i) enlarge the preference pool with varied negative examples, (ii) harmonize training and inference prompts, and (iii) run ORPO as a post-SFT refinement step (Hong, Lee, and Thorne Reference Hong, Lee, Thorne, Salakhutdinov, Kolter and Heller2024; Ouyang et al. Reference Ouyang, Jeff, Jiang, Alice, Naumann, Globerson, Saenko, Hardt and Levine2022). The suboptimal performance of ORPO highlights the sensitivity of preference-based alignment methods to data quality and prompt design. Future iterations will focus not just on enlarging the preference pool, but also on diversifying it. This includes generating “better” negative examples that are subtly incorrect (e.g., grammatically fluent but semantically inaccurate) rather than obviously flawed. This would force the model to learn finer-grained distinctions. Beyond ORPO, exploring other alignment techniques like Kahneman–Tversky optimization (KTO), which does not require paired preference data, could offer a more robust solution for low-resource scenarios. A comparative study of these different alignment methods could yield a clearer understanding of which strategies are best suited for the unique challenges of historical texts, where human preferences might involve balancing literal accuracy with historical authenticity.

Adaptation to other historical languages

The presented workflow, comprising metadata-rich corpus curation, custom tokenizer training and two-stage alignment with continuous domain expert oversight, offers a transferable methodological blueprint for other low-resource historical languages (e.g., Middle High German, early modern Spanish and regional Malay manuscripts). Although the current investigation focuses specifically on Early Modern Dutch legal texts, the core components of this framework, particularly the integration of domain expertise throughout the pipeline, provide a reproducible approach adaptable to diverse historical contexts and text types.

It is acknowledged that the present implementation utilizes a specialized legal corpus; consequently, future applications would benefit from validating the approach across multiple genres and domains. The essential transferable elements include the PEFT strategies, evaluation protocols combining automatic metrics with expert review and tokenizer adaptation methodologies. However, successful application to new languages necessitates language-specific resources, particularly curated parallel corpora developed in collaboration with subject matter experts.

By emphasizing verifiable translation through sustained domain expert involvement, this pipeline model could help unlock extensive archives for linguists, historians and the broader public. The framework’s ultimate adaptability will depend on addressing language-specific challenges while maintaining the core principle that historical translation requires continuous scholarly validation to ensure both accuracy and preservation of documentary authenticity.

Interactive and verifiable translation

Future systems should move beyond static translation and toward an interactive, human-in-the-loop framework. This could involve developing an interface where historians can not only correct translations but also see the model’s confidence scores for specific words or phrases. The system could offer multiple translation options for ambiguous terms, annotated with explanations drawn from historical dictionaries or linguistic corpora. Such a tool would transform the model from a “black box” into a transparent research assistant. Furthermore, integrating a mechanism for the model to learn from these expert corrections in real-time (online learning) would create a continuously improving system that becomes more attuned to the specific needs and nuances of a particular research project or archive.

Interdisciplinary collaboration

This research highlights the critical importance of sustained interdisciplinary collaboration between computational methods and humanities scholarship. Future work should develop structured frameworks for domain experts and technical researchers to co-design translation systems that balance technical efficiency with historical accuracy. Such partnerships are essential for addressing epistemological questions about how MT shapes historical interpretation while ensuring tools enhance rather than replace traditional scholarly practices. Establishing sustainable collaboration models will be crucial for adapting this approach to diverse historical contexts and text types.