OLDI Seed Corpus Overview

• OLDI Seed Corpus is a multilingual seed dataset combining professionally translated, curated Wikipedia content to kickstart low-resource MT.
• It uses a pivot-based approach where the French partition acts as a strategic resource for tri-parallel corpus construction and regional language support.
• Developed through a rigorous MT-plus-post-editing pipeline with nine systems, the corpus balances technical terminology with inherent source noise.

Searching arXiv for papers on the OLDI Seed Corpus and closely related seed-corpus expansion work. The OLDI Seed Corpus is a small, high-quality, professionally translated multilingual dataset conceived as a “seed” for bootstrapping machine translation in low-resource languages, especially when parallel data is scarce. It was initially created under the No Language Left Behind effort and later adopted by the Open Language Data Initiative (OLDI), where it functions as an open foundational resource for data creation, benchmarking, and method development in low-resource MT (Marmonier et al., 4 Aug 2025). The source material consists of approximately 6,000 English sentences sampled from a curated list of core Wikipedia articles to ensure broad topic coverage across encyclopedic domains, while also inheriting stylistic irregularities and extraction noise characteristic of Wikipedia segments (Marmonier et al., 4 Aug 2025). This dual character—high-value encyclopedic content combined with user-generated-content artifacts—makes the corpus simultaneously useful for translation research and challenging as a translation benchmark (Marmonier et al., 4 Aug 2025).

1. Definition, origin, and scope

The defining purpose of the OLDI Seed Corpus is to provide a “small but high-quality, professionally translated dataset” for dozens of low-resource languages, so that translation systems can be kickstarted even when conventional parallel corpora are unavailable or extremely limited (Marmonier et al., 4 Aug 2025). In the WMT 2025 OLDI shared task framing, it serves as a reusable seed resource that the research community can extend to additional languages (Marmonier et al., 4 Aug 2025).

The source side is English, and the data were sampled from a curated set of core Wikipedia articles chosen for broad encyclopedic coverage (Marmonier et al., 4 Aug 2025). The English source partition used in the French work comprises 6,193 segments and 136,656 source-side words, denoted as $N_{\text{sent}} = 6{,}193$ and $N_{\text{tok\_source}} = 136{,}656$ (Marmonier et al., 4 Aug 2025). The corpus therefore occupies an intermediate position between tiny handcrafted evaluation sets and large noisy crawled corpora: it is intentionally limited in size, but designed to have enough breadth to support terminology-rich translation pipelines, including backtranslation (Marmonier et al., 4 Aug 2025).

A central property of the resource is that it is not intended as a terminal product. The French partition paper explicitly states that such partitions are meant to function as pivot resources and as inputs to downstream corpus construction, rather than as isolated end-products (Marmonier et al., 4 Aug 2025). This suggests that “seed” should be understood operationally: the corpus is valuable not only for direct model training, but for initiating broader multilingual data ecosystems.

2. Source composition and linguistic characteristics

The source material combines formal encyclopedic prose with user-generated-content artifacts from Wikipedia, producing a mix of technical terminology, stylistic irregularities, segmentation noise, and occasional extraction errors (Marmonier et al., 4 Aug 2025). The reported topical range includes subjects such as cartography, bionanotechnology, Gothic architecture, and Hilbert’s problems (Marmonier et al., 4 Aug 2025). This breadth is one of the corpus’s explicit design goals, since it is intended to support technical terminology across many domains (Marmonier et al., 4 Aug 2025).

At the same time, the corpus is not uniformly clean. The French partition paper documents typographical issues, garden-path sentences, incomplete segments, markup noise, and source ambiguities (Marmonier et al., 4 Aug 2025). Examples include malformed source sentences, incomplete segments such as “Another early globe, the Hunt–Lenox Globe, ca.”, and ambiguous or ungrammatical passages that required consultation and cross-lingual comparison during post-editing (Marmonier et al., 4 Aug 2025). Some source-side anomalies were preserved when OLDI guidelines required fidelity to the original segment, even when extraction issues were apparent (Marmonier et al., 4 Aug 2025).

These characteristics matter methodologically. A corpus that combines technical, encyclopedic terminology with noisy segmentation and occasional source corruption is not equivalent to a standard parallel benchmark built from edited prose. A plausible implication is that the OLDI Seed Corpus stresses translation systems along at least two axes simultaneously: domain-specific lexical precision and robustness to irregular source form. The French partition paper makes this challenge explicit by emphasizing the tension between literal adherence to OLDI guidelines and the need to correct source-side disfluencies inherited from Wikipedia (Marmonier et al., 4 Aug 2025).

3. The French partition as a pivot-language realization

The first French partition of the OLDI Seed Corpus was presented for the WMT 2025 shared task as a deliberate pivot resource for under-resourced regional languages of France, rather than as a response to French data scarcity per se (Marmonier et al., 4 Aug 2025). The rationale is sociolinguistic: translators for languages such as Francoprovençal, Occitan, and Picard are more likely to be native in French than in English, and technical or formal registers are mediated through French in practice (Marmonier et al., 4 Aug 2025).

The paper formalizes this pivot scenario using source language $S$ (English), pivot $F$ (French), and regional language $R$, with aligned assets $D_{SF}$ and $D_{FR}$ linked by segment IDs and source URLs (Marmonier et al., 4 Aug 2025). It then defines indirect mining of $D_{SR}$ through alignment on the French segment $f_i$ (Marmonier et al., 4 Aug 2025). This makes the French partition more than a bilingual dataset: it is infrastructure for tri-parallel corpus construction.

The reported data profile for the French work preserves the full 6,193-segment English source set (Marmonier et al., 4 Aug 2025). Post-editing statistics indicate that 3,043 segments required no edits to the selected MT hypothesis, corresponding to $$p_{\text{perfect}} = 3{,}043 / 6{,}193 \approx 0.4914$$, and that DeepSeek-R1 supplied the perfect translation in 2,503 instances, or $N_{\text{tok\_source}} = 136{,}656$0 (Marmonier et al., 4 Aug 2025). These figures characterize the difficulty of the corpus and the quality of the hypothesis pool used in human post-editing.

4. Creation workflow and quality control

The French partition was constructed through a multi-system MT-plus-post-editing pipeline involving nine MT systems: OPUS-MT en–fr, NLLB 3.3B, NLLB-200 600M distilled, MADLAD-400 3B, Llama 4 Scout under several prompting conditions, and DeepSeek-R1 (Marmonier et al., 4 Aug 2025). The traditional sequence-to-sequence systems used sentence-level beam search with beam size $N_{\text{tok\_source}} = 136{,}656$1 (Marmonier et al., 4 Aug 2025). For document-level LLM runs, segments were first grouped by source URL and ordered by numerical ID to reconstruct documents, after which the generated block translations were segmented and realigned to the original segments, with manual correction required for approximately 5% of documents (Marmonier et al., 4 Aug 2025).

Human post-editing was carried out by two native French speakers with C2-level English, with a native British English speaker with C2-level French consulted for source ambiguities (Marmonier et al., 4 Aug 2025). The interface was a custom-built Vue.js/Tailwind CSS tool that displayed all nine MT hypotheses per segment, sorted by COMET-Kiwi QE scores, so that editors could select and refine the most promising candidate (Marmonier et al., 4 Aug 2025). Post-editing targeted both fluency and accuracy: fluency improvements addressed source-side disfluencies and segmentation errors, while accuracy improvements required systematic external terminological research, since FranceTerme was often insufficient (Marmonier et al., 4 Aug 2025).

The final output was checked with Grammalecte for spelling and grammar, and the authors explicitly excluded commercial systems with restrictive terms of service such as Google and DeepL to ensure reusability of MT outputs (Marmonier et al., 4 Aug 2025). The final French partition is released under CC BY-SA 4.0, consistent with the source Seed corpus, and supplementary data include all nine MT hypotheses plus the human reference for each segment (Marmonier et al., 4 Aug 2025).

The following table summarizes the core structural properties reported for the French partition.

Property	Reported value	Source
Segments	6,193	(Marmonier et al., 4 Aug 2025)
Source-side words	136,656	(Marmonier et al., 4 Aug 2025)
Perfect MT hypotheses	3,043	(Marmonier et al., 4 Aug 2025)
DeepSeek-R1 perfect cases	2,503	(Marmonier et al., 4 Aug 2025)
License	CC BY-SA 4.0	(Marmonier et al., 4 Aug 2025)

5. Evaluation and empirical quality profile

The French partition paper evaluates the final corpus and all raw MT hypotheses using MetricX-24, specifically metricx-24-hybrid-xl-v2p6, a hybrid reference-based/reference-free metric trained on MQM and DA human judgments, where lower error is better (Marmonier et al., 4 Aug 2025). On the full dataset of 6,193 segments, the human post-edited translations achieved an average error of 2.0790 with a 95% confidence interval of [2.04, 2.12], forming Group A alone (Marmonier et al., 4 Aug 2025). The closest machine systems on the full dataset were NLLB-3.3B at 2.2223, MADLAD-400-3B at 2.2290, and Llama-4-Scout segment-level at 2.2437, all in Group B (Marmonier et al., 4 Aug 2025).

When excluding the 165 DeepSeek-R1 refusal cases, the human post-edited output remained the top system with 2.0871 [2.05, 2.12], while DeepSeek-R1 moved into the top machine tier at 2.2238 [2.18, 2.26], alongside NLLB-3.3B, MADLAD-400-3B, and Llama-4-Scout segment-level (Marmonier et al., 4 Aug 2025). The paper also reports that 76% of outputs in the Llama document-level guidelines ablation were identical, and for differing cases the average Translation Edit Rate was approximately 9.48, indicating mostly minor variations (Marmonier et al., 4 Aug 2025).

These results support two conclusions stated in the paper. First, human post-edited translations are significantly better than any single MT hypothesis (Marmonier et al., 4 Aug 2025). Second, segment-level prompting outperformed the tested document-level Llama conditions, and adding corresponding French Wikipedia article context degraded scores rather than improving them (Marmonier et al., 4 Aug 2025). This is important because it suggests that for this corpus, broader contextual prompting does not automatically compensate for noisy source segmentation or technical lexical demands.

6. Research uses and methodological extensions

The corpus is explicitly positioned for several uses: MT training and evaluation in the Wikipedia domain, terminology and domain-adaptation research, quality estimation and preference optimization, and pivot-based corpus construction (Marmonier et al., 4 Aug 2025). Because the supplementary package includes all nine MT hypotheses alongside the post-edited reference, it is also suitable for post-editing studies and QE model development (Marmonier et al., 4 Aug 2025).

Several later or related works suggest concrete ways in which an OLDI Seed Corpus can be extended once a high-quality seed parallel dataset exists. One method proposes augmenting a small seed parallel corpus without additional monolingual data by masking one content word at a time with a multilingual masked LLM such as XLM-R, generating contextual variants, and filtering source–target candidate pairs using LaBSE cosine similarity and optional TransQuest QE validation (Kumari et al., 2024). In the reported English–Hindi illustration, masking 5 English tokens and 6 Hindi tokens with top-$N_{\text{tok\_source}} = 136{,}656$2 generated 3,000 candidate pairs, from which 200 were retained with LaBSE cosine similarity $N_{\text{tok\_source}} = 136{,}656$3, and all selected pairs also had TransQuest scores $N_{\text{tok\_source}} = 136{,}656$4 (Kumari et al., 2024). This suggests a plausible pathway for OLDI-style corpus multiplication from a trusted seed set.

For domain-specific corpus selection from large background text, seed-driven filtering has also been used in ASR LLM adaptation. A user-provided glossary can be expanded through shallow morphology and word2vec neighbors, then used to extract an in-domain subset from a large generic corpus by retaining documents containing at least one seed word not already in the baseline lexicon (Gretter et al., 2021). The authors explicitly describe the selected subset as an “OLDI Seed Corpus” for the domain (Gretter et al., 2021). This suggests that, beyond bilingual translation data, the OLDI concept can extend to seed-guided domain corpus construction more broadly.

For web-scale monolingual corpus expansion, SAUCE represents another compatible direction: starting from a small seed corpus of relevant documents, it constructs truncated sparse document signatures and retrieves a large focused corpus with high within-domain lexical coverage (Wahed et al., 2021). Reported results on approximately 200 million Common Crawl documents show 93.35% coverage in Astronomy with a 6-minute runtime, compared with 92.54% for RoBERTa in 27 minutes (Wahed et al., 2021). A plausible implication is that OLDI-style seed resources can function as high-precision anchors not only for direct translation but also for scalable retrieval and expansion pipelines.

7. Limitations, controversies, and broader significance

The main limitations reported for the French partition are the tension between literal adherence to OLDI guidelines and correction of Wikipedia-derived disfluencies, the persistence of residual source artifacts despite careful post-editing, and the fact that the French dataset is only a first step because it does not yet include regional-language pairs (Marmonier et al., 4 Aug 2025). More broadly, the source data’s encyclopedic domain bias and uneven quality introduce both value and noise: the corpus is terminology-rich and broadly topical, but also inherits ambiguities and segmentation defects from Wikipedia extraction (Marmonier et al., 4 Aug 2025).

A potential misconception is that the OLDI Seed Corpus is simply a small benchmark set. The French partition paper frames it differently: as a strategic pivot resource and an open seed that can be extended into richer multilingual assets (Marmonier et al., 4 Aug 2025). Another misconception is that a high-resource pivot such as French would be unnecessary in a low-resource initiative. The reported rationale is explicitly sociolinguistic and practical: French better aligns with translator availability and terminological practice for the regional languages of France (Marmonier et al., 4 Aug 2025).

In the broader research landscape, the OLDI Seed Corpus exemplifies a recurring design principle: small, curated seed resources can be leveraged disproportionately through expansion, retrieval, augmentation, and pivoting. Methods for masked-LM-based parallel augmentation (Kumari et al., 2024), seed-guided domain text selection (Gretter et al., 2021), and web-scale corpus expansion from small expert seed sets (Wahed et al., 2021) all reinforce that principle. The OLDI Seed Corpus therefore occupies a distinctive position: it is simultaneously a multilingual dataset, a translation benchmark, a pivot layer for low-resource data creation, and a prototype for seed-centered corpus engineering in multilingual NLP.