GaMS-9B-Instruct: English-Slovene MT LLM

• GaMS-9B-Instruct is an open-source, 9B-parameter decoder-only large language model designed for instruction tasks with a unique Slovene adaptation.
• It leverages synthetic preference data and Direct Preference Optimization (DPO) to enhance performance in long-document English-to-Slovene machine translation.
• The model demonstrates significant error reduction and improved translation quality, offering a scalable solution for low-resource language MT.

Searching arXiv for the primary paper and closely related methodological context. GaMS-9B-Instruct is an open-source, decoder-only LLM with about 9 billion parameters whose documented significance lies in its role as a Slovene-capable instruction model and, in particular, as the base system adapted for high-reliability English→Slovene machine translation in “Improving LLMs for Machine Translation Using Synthetic Preference Data” (Vajda et al., 20 Aug 2025). In that work, GaMS-9B-Instruct is both the starting point and the main beneficiary of a targeted alignment pipeline based on synthetic preference data and Direct Preference Optimization (DPO), transforming a general-purpose instruction-tuned model into a substantially more reliable long-document translator for Slovene.

1. Model identity and linguistic adaptation

GaMS-9B-Instruct is based on Google’s Gemma 2 architecture, specifically a transformer decoder-only LLM, and is derived from an English-centric base model that was subsequently adapted through continued pretraining on a sizeable Slovene corpus (Vajda et al., 20 Aug 2025). Two implementation choices are central to that adaptation: a Slovene-adapted subword tokenizer and embedding-initialization techniques that transfer knowledge from the original English embeddings to a Slovene-focused vocabulary.

The model was instruction-tuned for general-purpose tasks, including translation, but the paper characterizes that translation capability as mostly sentence-level and trained with relatively limited Slovene data compared to English (Vajda et al., 20 Aug 2025). This distinction is important: GaMS-9B-Instruct is not presented as a model originally specialized for long-form translation workflows, but as a general instruction model whose Slovene competence makes it a strong candidate for post hoc MT alignment.

Within the paper’s framing, GaMS-9B-Instruct occupies a specific niche among open models for English→Slovene translation. According to SloBench, it is one of the best open-source English→Slovene translators, yet it exhibits recurrent reliability failures on long inputs: wrong-language outputs, truncation, and formatting artifacts such as prepended strings like “Slovenski prevod:” (Vajda et al., 20 Aug 2025). A plausible implication is that the model’s preexisting linguistic competence was already substantial, while its output policy for document-level MT remained underconstrained.

2. Position in English→Slovene machine translation

The paper situates GaMS-9B-Instruct in a low-resource MT setting where Slovene has minimal high-quality parallel corpora and little presence in large multilingual pretraining datasets (Vajda et al., 20 Aug 2025). This setting makes open long-document translation especially difficult: existing open-source MT systems for Slovene, including RSDO encoder–decoder MT, are described as performing poorly out-of-domain, while the strongest systems for Slovene—DeepL, GPT, Claude, and Gemini—are proprietary and expensive for large-scale corpus translation.

A central comparison in the study is between GaMS-9B-Instruct and EuroLLM-9B-Instruct. EuroLLM-9B-Instruct is described as another multilingual 9B model with only approximately 1% Slovene in its training data; it performs well and reliably on many MT tasks but is not Slovene-specialized (Vajda et al., 20 Aug 2025). Conceptually, the paper treats EuroLLM-9B-Instruct as a strong multilingual generalist and GaMS-9B-Instruct as a Slovene-oriented specialist whose long-document translation behavior is insufficiently optimized.

That distinction determines the paper’s design. EuroLLM-9B-Instruct is not the target of alignment; rather, it functions as a source of alternative translations and as a “second opinion” during synthetic preference construction. GaMS-9B-Instruct is the model being aligned via DPO (Vajda et al., 20 Aug 2025). This is methodologically significant because the preference pipeline does not require human rankings or a large Slovene parallel corpus; it requires two reasonably capable MT-capable LLMs whose outputs can be compared.

3. DPO specialization with synthetic preferences

The paper improves GaMS-9B-Instruct using Direct Preference Optimization instead of RLHF (Vajda et al., 20 Aug 2025). The preference dataset is written as

$\mathcal{D}=\{(x,y_w,y_l)\},$

where $x$ is the input, $y_w$ is the chosen translation, and $y_l$ is the rejected translation. DPO optimizes a policy $\pi_\theta$ against a reference policy $\pi_{\text{ref}}$, here the original GaMS-9B-Instruct, with a loss built from the relative log-probabilities assigned to preferred and rejected outputs, modulated by the temperature parameter $\beta$ (Vajda et al., 20 Aug 2025).

The paper’s interpretation of DPO is pragmatic. It states that DPO is mathematically equivalent in objective to policy-optimization RLHF but avoids an explicit reward model and unstable PPO-style optimization (Vajda et al., 20 Aug 2025). In a synthetic-preference setting, that matters because adding a synthetic reward model would introduce another layer of possible noise. The reference policy also regularizes training, pushing the fine-tuned model toward preference-consistent behavior without excessive deviation from its original distribution.

The preferences encode both hard and soft quality signals. Hard failures include wrong-language outputs, truncation, and formatting artifacts; softer differences in fluency, adequacy, and style are captured by COMET-based ranking (Vajda et al., 20 Aug 2025). This mixture is central to the specialization of GaMS-9B-Instruct: the alignment target is not only higher average MT quality, but also structurally reliable translation behavior on long documents.

4. Preference dataset construction and training regime

The synthetic preference dataset begins with 67,000 English Wikipedia articles, about 26 million words, from the zIDsi Wikipedia-Markdown dataset, plus 30,000 English CC-News articles from Common Crawl (Vajda et al., 20 Aug 2025). For each article, the authors generate two candidate Slovene translations: one with GaMS-9B-Instruct using the Slovene prompt “Prevedi naslednje angleško besedilo v slovenščino.” and one with EuroLLM-9B-Instruct using the English prompt “Translate the following English text to Slovenian.” This yields approximately 107,000 dual-translation pairs.

Three heuristic filters identify hard failures. Wrong-language outputs are detected with FastText’s pretrained language-ID classifier; if one output is confidently Slovene and the other is classified as a different language, the Slovene output is preferred (Vajda et al., 20 Aug 2025). Truncation is detected by comparing translation length to source length; if the translation is under 50% of source length, it is deemed truncated. Formatting artifacts are targeted through synthetic pairs in which a clean translation is preferred over the same translation with an undesirable prefix such as “Slovenski prevod:” prepended.

For subtler quality differences, the paper uses the reference-less COMETKiwi DA model Unbabel/wmt22-cometkiwi-da (Vajda et al., 20 Aug 2025). A pair is created only if the absolute score difference satisfies

$|\mathrm{COMET}(y_1)-\mathrm{COMET}(y_2)|>0.05.$

The higher-scoring output becomes the chosen response. This threshold is explicitly intended to avoid treating small COMET fluctuations as genuine preferences.

The final dataset contains approximately 35,000 preference pairs: about 25,000 from Wikipedia and about 10,000 from CC-News (Vajda et al., 20 Aug 2025). Their distribution is 22% incorrect-language pairs, 3% truncation pairs, 20% synthetic formatting pairs, and 55% COMET-based quality pairs. The paper therefore combines negative constraints on unacceptable outputs with relative judgments about translation quality.

Training is performed on the Slovene HPC Vega supercomputer using HuggingFace Transformers, TRL, Accelerate, DeepSpeed, and PEFT with LoRA rank 64 (Vajda et al., 20 Aug 2025). The setup uses 4 compute nodes, with 4 NVIDIA A100 40GB GPUs per node, for a total of 16 A100 GPUs; NVLINK provides 600 GB/s aggregate within node, and inter-node communication uses dual 200 Gb/s InfiniBand in a Dragonfly+ topology. Memory efficiency is achieved with ZeRO Stage 2 and gradient checkpointing. Hyperparameter search on 24,000 Wikipedia preference pairs with a 1,000-pair validation set explores $\beta \in \{0.1,0.2\}$ and learning rates in $\{1\times10^{-6},4\times10^{-7},1\times10^{-7}\}$ over 3 epochs, with micro batch size 1 per GPU, global batch size 16, 1500 warmup steps, and a cosine_with_min_lr scheduler (Vajda et al., 20 Aug 2025). The best configuration is $x$0 and $x$1, with validation loss approximately 0.315 on Wikipedia-only preferences; retraining on Wikipedia plus CC-News with the same hyperparameters yields a lower validation loss of approximately 0.255. The resulting model is published as GaMS-9B-DPO-Translator.

5. Empirical behavior before and after alignment

On the SloBench English→Slovene MT benchmark, the paper reports that GaMS-9B-DPO-Translator improves over GaMS-9B-Instruct across BERTScore, BLEU, METEOR, and chrF (Vajda et al., 20 Aug 2025). For the baseline GaMS-9B-Instruct, the reported scores are BERTScore 0.8713, BLEU (avg) 0.2773, METEOR (avg) 0.5616, chrF (avg) 0.5928, BLEU (corpus) 0.3209, and chrF (corpus) 0.5928. For GaMS-9B-DPO-Translator, they are 0.8726, 0.2810, 0.5663, 0.5967, 0.3252, and 0.5967 respectively. EuroLLM-9B-Instruct attains 0.8741 BERTScore, 0.2927 BLEU (avg), 0.5792 METEOR (avg), 0.6055 chrF (avg), 0.3273 BLEU (corpus), and 0.6055 chrF (corpus), while GaMS-27B-Instruct records 0.8734, 0.2866, 0.5688, 0.5986, 0.3246, and 0.5986. The paper explicitly notes that GaMS-9B-DPO-Translator is comparable to GaMS-27B-Instruct despite using only 9B parameters.

The long-document evaluation is more distinctive. On 500 held-out Wikipedia articles and 500 held-out CC-News articles, each translated by baseline GaMS-9B-Instruct, GaMS-9B-DPO-Translator, and EuroLLM-9B-Instruct, the paper measures hard error rates using FastText language ID and the same truncation criterion of output length below 50% of source length (Vajda et al., 20 Aug 2025). EuroLLM-9B-Instruct shows 1.0% language error, 0.4% truncation error, and 1.4% combined error. Baseline GaMS-9B-Instruct shows 9.5%, 3.5%, and 13.0%. GaMS-9B-DPO-Translator shows 0.6%, 0.2%, and 0.8%.

For cases where none of the three models exhibits trivial errors, COMETKiwi DA averages are also reported (Vajda et al., 20 Aug 2025). EuroLLM-9B-Instruct receives 0.727 on Wikipedia, 0.667 on CC-News, and 0.695 overall. Baseline GaMS-9B-Instruct receives 0.722, 0.680, and 0.698. GaMS-9B-DPO-Translator receives 0.757, 0.715, and 0.735. The paper summarizes this as a COMET score gain of around 0.04 and 0.02, respectively, over the two baseline models on translating Wikipedia articles, and more broadly presents gains in the range of roughly 0.02–0.04 depending on subset and rounding.

Qualitatively, the paper characterizes the aligned model as more fluent and complete, especially for long documents, and as more consistent in avoiding language mixing, incomplete translations, and meta prefixes (Vajda et al., 20 Aug 2025). Although the paper does not reproduce extended example translations, the reported reduction from 13.0% to 0.8% combined hard-error rate is the clearest empirical expression of that shift in behavior.

6. Interpretation, limitations, and broader relevance

The paper’s central lesson is not merely that GaMS-9B-Instruct can be improved, but that a medium-sized instruction-tuned LLM can be specialized for MT using relatively few, easily produced data resources and no human-labeled preferences (Vajda et al., 20 Aug 2025). The data creation cost is reported as about 12 GPU-hours, and fine-tuning as about 112 GPU-hours; the minimum viable hardware is stated as one node with 4 A100s and under 28 hours of training thanks to ZeRO Stage 2. This suggests an operationally accessible path for low-resource-language MT adaptation.

At the same time, the paper is explicit about several limitations. It does not include ablations such as COMET-only versus heuristics-only preference construction, and it does not empirically compare DPO against full RLHF or supervised MT fine-tuning (Vajda et al., 20 Aug 2025). The claimed language-agnostic character of the method depends on the existence of at least two reasonably capable translators for the target language pair. The paper also notes risks of metric bias from COMETKiwi, dependence on the quality of the generating LLMs, and possible Wikipedia-style bias from training primarily on encyclopedic material.

These cautions matter for interpreting GaMS-9B-Instruct itself. The documented improvements are specific to a targeted English→Slovene MT alignment pipeline rather than a general claim about all downstream tasks. What the evidence supports is narrower and technically clearer: GaMS-9B-Instruct is a strong Slovene-capable general LLM whose long-document translation policy can be substantially improved through synthetic preference data, heuristically encoded reliability constraints, and DPO regularized against the original model (Vajda et al., 20 Aug 2025). Future directions named in the paper include scaling the procedure to GaMS-27B-Instruct, adding more conversational, legal, and news texts, using curriculum DPO based on preference difficulty, and exploring other preference methods such as GRPO.