TranslatePsy-AfriSLM: Optimized Machine Translation for African Languages

📄 Paper (arXiv) · 💻 GitHub · 🤗 Hugging Face Collection

The rapid progress in Artificial Intelligence has largely bypassed African languages, creating a digital divide that limits AI adoption for over a billion people across the continent [1]. For many African communities, language technology, computational resources and robust connectivity remain fundamental barriers to AI access. There is only limited support for local African languages and unreliable connectivity to (expensive) cloud-based AI systems. Bridging Africa’s AI divide therefore requires not only better language support, but affordable AI that can operate on personal devices where connectivity is limited.

We introduce TranslatePsy-AfriSLM, an open-source suite of machine translation resources covering 19 Sub-Saharan African languages. Our research aims to (via multiple releases) provide the language-access layer for bringing AI-powered knowledge to African communities in their own languages, directly on everyday devices and without continuous internet connectivity. The first release of TranslatePsy-AfriSLM is documented in our forthcoming EMNLP 2026 paper, titled 'High-Quality Data Scaling for Low-Resource Machine Translation', where we demonstrate how our rigorous quality-estimation filtering and synthetic data generation allows SLMs with as few as 0.8B parameters to rival, and even outperform, frontier open-source LLMs and specialised translation models on African languages.

This points towards a broader vision for AI in Africa such as delivering educational and scientific content to students, providing agricultural guidance to farmers, supporting multilingual communication in areas with unreliable connectivity and other large-scale AI applications. For example, translation tools could extend access to critical domains such as healthcare, when paired with models like Tether’s QVAC MedPsy, a medical and healthcare foundation model. This would enable health information to be delivered and understood in local African languages, subject to the rigor and safeguards required for medical applications. Our efforts ultimately lead to applications that are not constrained by dominant languages, cloud connectivity, or expensive infrastructure, but are built for African languages, deployable on the devices communities already own, capable of bringing essential knowledge and AI-powered services to people wherever they are.

Figure 1: SSA-COMET on Flores-200, BOUQuET, and Smol. Even our smallest model, TranslatePsy-AfriSLM-0.8B, matches or outperforms general-purpose, dedicated, and African-specialised systems up to 152× its size (e.g. Qwen3.5-122B-A10B, TranslateGemma-27B) on all three benchmarks.

🚀 Key Highlights

• SLMs Beat Frontier Behemoths: Our TranslatePsy-AfriSLM-0.8B outperforms much larger systems, including Qwen3.5-122B-A10B, TranslateGemma-27B, and NLLB-3.3B across major African machine translation benchmarks.
• 19 African Languages: TranslatePsy-AfriSLM covers 19 African languages including languages that remain relatively underrepresented in existing AI machine translation models: Afrikaans, Amharic, Hausa, Igbo, Kinyarwanda, Lingala, Luganda, Malagasy, Nyanja, Oromo, Shona, Somali, Southern Sotho, Swahili, Tswana, Wolof, Xhosa, Yoruba, and Zulu.
• Up to 96% Fewer Training Tokens: Our unified quality estimation (QE) filtering filters noisy and redundant training data, reducing the open-source training budget from 44.93B to 1.76B tokens, a 96% reduction without degrading translation performance.
• Synthetic Data Superiority: Our high-quality synthetic parallel data dominates open-source corpora across the quality-efficiency Pareto frontier showing that data quality and efficiency matters more than simply scaling data volume.
• Conversational Translation: TranslatePsy-AfriSLM retains multi-turn translation, language identification, and cross-lingual conversational capabilities, going beyond rigid sentence-level translation systems.
• Run on a laptop or smartphone: We additionally share GGUF models for llama.cpp, which are quantized models for on-device deployment, providing further efficiency and convenience. No need for expensive GPUs, internet access or advanced hardware!
• Open-Source Release: Tether Data, S.A. de C.V. (Tether AI Research, we, us, our) releases the model weights, inference code and data-processing scripts under the Apache 2.0 license for research and educational purposes. The synthetic training data is released under CC-BY-NC 4.0 (Creative Commons Attribution–NonCommercial 4.0) for research and educational purposes.

Copyright Complaints: We will take appropriate actions in response to notice of copyright infringement. If you believe your work has been used or copied in a manner that infringes upon your intellectual property rights, please email data-apps@tether.io identifying and describing both the copyrighted work and alleged infringing content to file a notice of infringement.

---

🌍 Motivation: Why African MT Is a Data Problem

Machine translation is critical for communication, education, and cross-border collaboration across Africa, yet even large open-source LLMs remain relatively weak in many African languages. Adapting smaller models is promising, but the main bottleneck is high-quality training data at sufficient scale.

Existing African MT resources face a persistent trade-off between scale, coverage, and quality:

• Large but noisy corpora: OPUS, MALA, WMT22, and Fine Translations provide millions of parallel sentences, but contain substantial duplication, imbalance, and noisy translations.
• Curated datasets: AfriNLLB offers a cleaner supervision, but covers fewer languages and remains relatively limited in scale.
• Human-quality datasets: SMOL, AfriDOC-MT, MMT-Africa, and LAFAND-MT provide strong training signals per example, but are too small to drive large-scale SLM adaptation on their own.

This leads to our central question: How can we scale our training data budget while maintaining high quality?

🛠️ Methodology: Building a High-Quality Training Pipeline

TranslatePsy-AfriSLM is built around a simple idea: scale high-quality training signals rather than just raw volume.  Our data curation pipeline starts from large-scale open-source and synthetic parallel data, then progressively cleans, scores, and filters it to retain the most useful training signal for African language machine translation.

Figure 2: The TranslatePsy-AfriSLM data pipeline. Raw parallel and monolingual sources are cleaned, expanded with NLLB-3.3B synthetic generation, and filtered by a Unified QE module before data-quantity selection produces the final training mixtures.

1. Data Sourcing & Preprocessing

We build two large parallel-data pools:

• Open-source data: We collect 427M raw sentence pairs across 19 African languages from WMT22 [1], MALA [11], OPUS [10], and Fine Translations [12]. NOTE: this data mixture was used for research purposes only and will not be publicly released, however, we are releasing our synthetic data mixture listed below (our most performant training dataset).
• Synthetic data: starting from MADLAD-400 [13] monolingual text, we generate English ↔ African language parallel data using NLLB-3.3B via CTranslate2 [9].

Cleaning & Decontamination: Both undergo language-ID filtering, document and sentence cleaning, exact/fuzzy deduplication, and strict decontamination against all evaluation sets.

2. Unified Quality Estimation (QE)

Quality estimators are models that have been trained to evaluate the quality of a translation. We found that no single QE estimator is consistently optimal across evaluation metrics. In our ablations, data selected by each estimator tends to perform particularly well under its corresponding evaluation metric. However, in order to obtain a more robust filtering signal, we combine AfriCOMET [8], SSA-COMET [4], and MetricX-24 [3] into a unified score.

Since these estimators differ in both scale and polarity, we calibrate their raw scores against approximately 352K human-translated sentence pairs using a robust z-score based on the median and median absolute deviation (MAD):

$${\mathrm{M}\mathrm{A}\mathrm{D}}_{d,m}=\mathrm{median}\left(\mid x_i-{\tilde{x}}_{d,m}\mid \right)\backslash mathrm\{MAD\}\_\{d,m\}\; =\; \backslash operatorname\{median\}\backslash left(\; \backslash left|x\_i\; -\; \backslash tilde\{x\}\_\{d,m\}\backslash right|\; \backslash right)$$

$$z_{i,d,m}=s_m\cdot \frac{0.6745(x_{i,d,m}-{\tilde{x}}_{d,m})}{{\mathrm{M}\mathrm{A}\mathrm{D}}_{d,m}}z\_\{i,d,m\}\; =\; s\_m\; \backslash cdot\; \backslash frac\{\; 0.6745\; \backslash left(x\_\{i,d,m\}\; -\; \backslash tilde\{x\}\_\{d,m\}\backslash right)\; \}\{\; \backslash mathrm\{MAD\}\_\{d,m\}\; \}$$

$${\bar{z}}_{i,d}=\frac{1}{3}\sum _{m=1}^3{z}_{i,d,m}\backslash bar\{z\}\_\{i,d\}\; =\; \backslash frac\{1\}\{3\}\; \backslash sum\_\{m=1\}^\{3\}\; z\_\{i,d,m\}$$

where $s_m\in \{+1,-1\}s\_m\; \backslash in\; \backslash \{+1,\; -1\backslash \}$ flips lower-is-better metrics so that higher normalized scores always indicate better estimated translation quality.

Scoring direction matters. QE filtering works best when scoring is aligned with the final training direction (X $\to \backslash rightarrow$ Y). Reversing the direction (Y $\to \backslash rightarrow$ X) leads to average downstream degradation of 12.0% on MetricX and 3.1% on SSA-COMET. This means that even for back-translated data, QE should be computed in the same direction as the training example.

3. Data Quantity Selection Strategies

Once each sentence pair receives a unified QE score, we study how to convert quality estimates into an effective training mixture. We compare three complementary strategies:

1. Thresholding: Keep only sentence pairs above a minimum $\bar{z}\backslash bar\{z\}$-score.
2. TopN Capping: Limit the number of examples per language pair to prevent higher-resource languages from dominating the mixture.
3. Bidirectional Expansion: Expand selected pairs into both translation directions while keeping QE scoring aligned with each final training direction.

These strategies allow us to study the effects of quality, language balance, and translation direction. We sweep them across open-source, synthetic, and combined data pools; the resulting quality-quantity trade-offs determine our final TranslatePsy-AfriSLM Synthetic Mix, discussed in the Results section.

4. Final Training Mix for SFT

Our scaling experiments select filtered synthetic data with thresholding and bidirectional expansion as the core TranslatePsy-AfriSLM Synthetic Mix. The final configuration uses a unified QE threshold of $\bar{z}\ge \backslash bar\{z\}\; \backslash ge$ 0.68, contains 32.37B training tokens, and reaches 0.632 SSA-COMET on BOUQuET (2B model). We then combine this MT dataset with two auxiliary mixtures that serve complementary roles:

• Instruct Mix (4.6M samples): approximately half consists of African-language instruction data (~2.3M examples, which are concatenated from multiple open-source African instruct datasets), complemented by ~2.3M general instruction-following examples (concatenated from two open-source datasets, see paper for full details). This is to ensure the models retain their instruction-following capabilities.
• Asia-Europe Mix (24M samples across 38 languages): High-resource parallel pairs included specifically to prevent catastrophic forgetting on non-African languages using a subset of OPUS-100 data. This is to ensure that major European and Asian languages are well supported. You can reproduce these datasets using our Github code.

Using this final mixture, we fine-tune Qwen3.5-0.8B, 2B, and 4B for one epoch with SFT, computing loss only on assistant target tokens. Training examples follow the Qwen3.5 chat format, with sequences capped at 2,048 tokens and packed for efficient training.

---

📊 Evaluation: Datasets & Metrics

We evaluate TranslatePsy-AfriSLM on 19 in-distribution African languages (Africa-IID) and 8 held-out African languages (Africa-OOD) across three complementary MT benchmarks.

🌐 Evaluation languages (Africa-IID & Africa-OOD)

📏 Benchmarks & metrics

The three benchmarks cover all 19 Africa-IID languages, while Africa-OOD coverage varies by benchmark. We use multiple neural and lexical metrics to evaluate translation quality from complementary perspectives (see table above).

---

🏆 Results & Insights

1. Data Scaling & Token Efficiency

We compare several data quantity selection strategies by fine-tuning Qwen3.5-2B with the corresponding data mix.

Quality filtering substantially improves token efficiency. For open-source data, filtering reduces the training budget from 44.93B to 1.76B tokens, a 96% reduction, while maintaining comparable SSA-COMET performance (see Figure 3 below).

Synthetic data scales even more effectively: filtered synthetic mixtures outperform open-source mixtures across nearly all training-token budgets. At larger scales, simply adding lower-quality open-source data provides little benefit and can even dilute the training signal.

Based on these results, we select a filtered synthetic mixture with bidirectional expansion as our final TranslatePsy-AfriSLM training data, using 32.37B tokens and reaching 0.632 SSA-COMET.

Figure 3: Quality–efficiency Pareto frontier — SSA-COMET vs. training tokens (log scale) on BOUQuET. Marker shape encodes the selection strategy and fill shade the QE threshold. Filtered synthetic (green) and synthetic + open-source (blue) mixtures dominate raw open-source data (red); the final TranslatePsy-AfriSLM sits at the top-right, far above the unfiltered open-source baseline.

2. Top Benchmark Performance

Fine-tuned on our final data mixture, TranslatePsy-AfriSLM demonstrates that model scale alone is not sufficient for strong African machine translation. Of the 19 Africa-IID languages, even our smallest 0.8B model outperforms Qwen3.5-122B-A10B and TranslateGemma-27B across Flores-200, BOUQuET, and Smol. It also matches NLLB-3.3B on Flores-200 and surpasses it on BOUQuET and Smol, despite using roughly one quarter of the parameters.

Scaling TranslatePsy-AfriSLM from 0.8B to 2B and 4B yields further consistent improvements, with the 4B model achieving the strongest SSA-COMET results across all three benchmarks among the evaluated systems. These results suggest that, for low-resource African MT, carefully curated post-training data can matter more than simply increasing model size.

Figure 4: The full TranslatePsy-AfriSLM family (0.8B, 2B, 4B) on SSA-COMET across Flores-200, BOUQuET, and Smol. Every size leads at its scale, and the 4B model achieves the strongest SSA-COMET overall, outperforming much larger general-purpose and specialised systems.

3. Generalization to Unseen African Languages

TranslatePsy-AfriSLM generalizes beyond the 19 languages seen during fine-tuning. On 8 held-out African languages, TranslatePsy-AfriSLM-2B improves over the Qwen3.5-2B backbone on all languages, with particularly strong gains for Sepedi, Bambara, and Akan.

These results suggest meaningful cross-lingual transfer beyond the training distribution (SSA-COMET scores), although gains are less uniform across other evaluation metrics.

Figure 5: Transfer to 8 held-out African languages never seen in fine-tuning. TranslatePsy-AfriSLM-2B improves SSA-COMET over its Qwen3.5-2B backbone on every language (both directions, averaged over Flores-200, BOUQuET, and Smol) — an average gain of +0.127 (+46%). Δ shows the per-language improvement.

4. Zero-Shot African-to-African Language Translation

Our training data contains only English ↔ African language pairs, meaning the models never observe direct African-to-African pairs during training. Despite this, TranslatePsy-AfriSLM-2B achieves the best aggregate performance across all four evaluation metrics on 20 zero-shot translation directions such as Swahili, Hausa, Yoruba, Zulu, and Amharic. This suggests that the improvements learned from English-centric post-training transfer to entirely unseen African-to-African translation directions.

Figure 6: Zero-shot African-to-African translation on 20 xx→yy directions (BOUQuET) among Swahili, Hausa, Yoruba, Zulu, and Amharic — none seen in training. Trained only on English↔African pairs, TranslatePsy-AfriSLM-2B still tops all four metrics against systems up to 61× its size. The MetricX-24 axis is reflected so taller is always better.

5. Mitigating Catastrophic Forgetting

Fine-tuning without the Asia-Europe Mix substantially degrades translation performance on non-African languages, with MetricX showing an 86.0% drop across the evaluated Asian and European languages (Figure 5). Adding the mix reduces this degradation to 10.3%, while preserving nearly identical gains on Africa-IID languages. We therefore include the Asia-Europe Mix to preserve broader multilingual capabilities.

Figure 7: Asia–Europe Mix ablation. Percent change vs. the Qwen3.5-2B baseline on BOUQuET for Asian, European, and Africa-IID languages (MetricX sign-flipped so higher is better everywhere). A light Asia–Europe mix (black) recovers most of the non-African forgetting seen without it (grey), averaging −39.0% → −4.1% across metrics, while African gains stay intact. Δ = mix − no-mix, in points.

6. Evaluation Robustness

Because our unified QE filter overlaps with several of the neural metrics that also serve as evaluation metrics, our reported gains could in principle reflect metric-specific bias rather than genuine translation quality. We address this concern from two independent angles: correlation with metrics excluded from filtering, and an independent LLM-as-a-judge evaluation.

Metric circularity. To test whether our rankings are an artifact of the filtering family, we compare our neural metrics (COMET-22, SSA-COMET, MetricX-24) against two lexical metrics that are never used for QE filtering: chrF++ and spBLEU. Across all system rankings, the neural and surface-level metrics move in tight alignment — every pairwise Spearman correlation is at least 0.95 in magnitude, and all are highly significant (p ≪ 0.001). Because the rankings hold under metrics outside the filtering family, our performance gains are unlikely to be an artifact of metric-specific optimization.

Table 1: Spearman rank correlation (ρ) between neural metrics (COMET-22, SSA-COMET, MetricX) and lexical metrics excluded from QE filtering (chrF++, spBLEU) across system rankings. All correlations are strong (|ρ| ≥ 0.95) and highly significant (p ≪ 0.001).

LLM-as-a-judge. As a second check, we conducted an LLM-as-a-judge evaluation comparing TranslatePsy-AfriSLM-2B against NLLB-3.3B and TranslateGemma-27B. We used two frontier models, GPT-5.5 and Claude Opus 4.8, as independent judges to provide semantic pairwise assessments of translation quality. For each language pair, we randomly sampled 100 sentences from the BOUQuET test set and consolidated the judge pairwise comparisons into sentence-level rankings. Aggregating how often each system was ranked first, second, or third, the resulting ordering (see Figure 8) agrees with our automatic-metric rankings, giving further evidence that the observed gains persist beyond the metrics used for data filtering.

Figure 8: LLM-as-a-judge rankings on BOUQuET (GPT-5.5 and Claude Opus 4.8 as judges). TranslatePsy-AfriSLM-2B is ranked #1 most often in both eng→xx and xx→eng, matching the automatic-metric ordering.

7. Conversational MT

Thanks to our Instruct mixes, TranslatePsy-AfriSLM models are not rigid input-output translators. They retain multi-turn translation, language identification, and cross-lingual conversational capabilities.

Figure 9: Multi-turn chat with TranslatePsy-AfriSLM-4B — sentence translation, language identification, and cross-lingual conversation across Lingala, Swahili, German, Amharic, Igbo, Chinese, Hausa, Somali, and Zulu in a single session.

🔮 Limitations & Future Work

While TranslatePsy-AfriSLM demonstrates the power of data scaling, several key challenges remain for future research:

1. Automated QE vs. Human Judgment: Reference-free QE models such as AfriCOMET, SSA-COMET and Metric-X share underlying model families with evaluation metrics. While synthetic data scores as high as human-translated pairs under automated metrics, absolute quality gaps compared to European languages still exist.
2. Compound Error Propagation: Large-scale automated curation pipelines carry the risk of compounding errors across data scraping, teacher model generation, QE filtering, and automated scoring. This warrants further investigation in future works.
3. Human Evaluation Campaigns: Future work must incorporate systematic human evaluation campaigns involving expert native linguists to validate fine-grained errors, dialectal nuances, and cultural context preservation.

Citation

@misc{gritta2026translatepsyafrislmhighqualitydatascaling,
      title={TranslatePsy-AfriSLM: High-Quality Data Scaling For Low-Resource Machine Translation},
      author={Milan Gritta and Patrik Lambert and Jihye Back and Amril Nazir},
      institution={Tether Data, S.A. de C.V. d.b.a. Tether AI Research},
      year={2026},
      eprint={2608.18655},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.18655},
}

📚 References

• [1] Adelani et al. (2022b): Findings of the WMT'22 shared task on large-scale machine translation evaluation for African languages. WMT.
• [2] Caswell et al. (2025): SMOL: Professionally translated parallel data for 115 under-represented languages.
• [3] Juraska et al. (2024): MetricX-24: The Google submission to the WMT 2024 metrics shared task. WMT.
• [4] Li et al. (2025): SSA-COMET: Do LLMs outperform learned metrics in evaluating MT for under-resourced African languages? EMNLP.
• [5] Moslem et al. (2026): AfriNLLB: Efficient translation models for African languages.
• [6] NLLB Team et al. (2022): No Language Left Behind: Scaling human-centered machine translation. arXiv:2207.04672.
• [7] Qwen Team (2026): Qwen3.5: Accelerating productivity with native multimodal agents.
• [8] Wang et al. (2024): AfriMTE and AfriCOMET: Enhancing COMET to embrace under-resourced African languages. NAACL.
• [9] Klein et al. (2020): CTranslate2: Fast inference engine for Transformer models. GitHub, OpenNMT.
• [10] Tiedemann (2012): Parallel data, tools and interfaces in OPUS. LREC.
• [11] Ji et al. (2024): EMMA-500: Enhancing massively multilingual adaptation of large language models (MALA corpus). arXiv:2409.17892.
• [12] Penedo et al. (2026): FineTranslations. Hugging Face Datasets.
• [13] Kudugunta et al. (2023): MADLAD-400: A multilingual and document-level large audited dataset.