We introduce Gemma 3, a multimodal addition to the Gemma family of lightweight open models, ranging in scale from 1 to 27 billion parameters. This version introduces vision understanding abilities, a wider coverage of languages and longer context - at least 128K tokens. We also change the architecture of the model to reduce the KV-cache memory that tends to explode with long context. This is achieved by increasing the ratio of local to global attention layers, and keeping the span on local attention short. The Gemma 3 models are trained with distillation and achieve superior performance to Gemma 2 for both pre-trained and instruction finetuned versions. In particular, our novel post-training recipe significantly improves the math, chat, instruction-following and multilingual abilities, making Gemma3-4B-IT competitive with Gemma2-27B-IT and Gemma3-27B-IT comparable to Gemini-1.5-Pro across benchmarks. We release all our models to the community.

While large language models (LLMs) have been increasingly adopted for machine translation (MT), their performance for specialist domains such as medicine and law remains an open challenge. Prior work has shown that LLMs can be domain-adapted at test-time by retrieving targeted few-shot demonstrations or terminologies for inclusion in the prompt. Meanwhile, for general-purpose LLM MT, recent studies have found some success in generating similarly useful domain knowledge from an LLM itself, prior to translation. Our work studies domain-adapted MT with LLMs through a careful prompting setup, finding that demonstrations consistently outperform terminology, and retrieval consistently outperforms generation. We find that generating demonstrations with weaker models can close the gap with larger model's zero-shot performance. Given the effectiveness of demonstrations, we perform detailed analyses to understand their value. We find that domain-specificity is particularly important, and that the popular multi-domain benchmark is testing adaptation to a particular writing style more so than to a specific domain.

We open-source SMOL (Set of Maximal Overall Leverage), a suite of training data to unlock translation for low-resource languages (LRLs). SMOL has been translated into 115 under-resourced languages, including many for which there exist no previous public resources, for a total of 6.1M translated tokens. SMOL comprises two sub-datasets, each carefully chosen for maximum impact given its size: SMOL-Sent, a set of sentences chosen for broad unique token coverage, and SMOL-Doc, a document-level source focusing on a broad topic coverage. They join the already released GATITOS for a trifecta of paragraph, sentence, and token-level content. We demonstrate that using SMOL to prompt or fine-tune Large Language Models yields robust ChrF improvements. In addition to translation, we provide factuality ratings and rationales for all documents in SMOL-Doc, yielding the first factuality datasets for most of these languages.

Data contamination -- the accidental consumption of evaluation examples within the pre-training data -- can undermine the validity of evaluation benchmarks. In this paper, we present a rigorous analysis of the effects of contamination on language models at 1B and 8B scales on the machine translation task. Starting from a carefully decontaminated train-test split, we systematically introduce contamination at various stages, scales, and data formats to isolate its effect and measure its impact on performance metrics. Our experiments reveal that contamination with both source and target substantially inflates BLEU scores, and this inflation is 2.5 times larger (up to 30 BLEU points) for 8B compared to 1B models. In contrast, source-only and target-only contamination generally produce smaller, less consistent over-estimations. Finally, we study how the temporal distribution and frequency of contaminated samples influence performance over-estimation across languages with varying degrees of data resources.

This paper investigates the impact of verbose LLM translations on evaluation. We first demonstrate the prevalence of this behavior across several LLM outputs drawn from the WMT 2024 general shared task on machine translation. We then identify the primary triggers of verbosity, including safety, copyright concerns, and insufficient context in short input queries. Finally, we show that ignoring this behavior unfairly penalizes more verbose LLMs according to both automatic and human evaluations, highlighting the need to address this issue for more accurate future evaluations.

A critical component in knowledge distillation is the means of coupling the teacher and student. The predominant sequence knowledge distillation method involves supervised learning of the student against teacher-decoded outputs, and is exemplified by the current state of the art, which incorporates minimum Bayes risk (MBR) decoding. In this paper we seek to integrate MBR more tightly in distillation training, specifically by using several high scoring MBR translations, rather than a single selected sequence, thus capturing a rich diversity of teacher outputs. Our experiments on English to German and English to Japanese translation show consistent improvements over strong baseline methods for both tasks and with varying model sizes. Additionally, we conduct a detailed analysis focusing on data efficiency and capacity curse aspects to elucidate MBR-n and explore its further potential.

While large language models (LLMs) often adopt finetuning to unlock their capabilities for downstream applications, our understanding on the inductive biases (especially the scaling properties) of different finetuning methods is still limited. To fill this gap, we conduct systematic experiments studying whether and how different scaling factors, including LLM model size, pretraining data size, new finetuning parameter size and finetuning data size, affect the finetuning performance. We consider two types of finetuning - full-model tuning (FMT) and parameter efficient tuning (PET, including prompt tuning and LoRA), and explore their scaling behaviors in the data-limited regime where the LLM model size substantially outweighs the finetuning data size. Based on two sets of pretrained bilingual LLMs from 1B to 16B and experiments on bilingual machine translation and multilingual summarization benchmarks, we find that 1) LLM finetuning follows a powerbased multiplicative joint scaling law between finetuning data size and each other scaling factor; 2) LLM finetuning benefits more from LLM model scaling than pretraining data scaling, and PET parameter scaling is generally ineffective; and 3) the optimal finetuning method is highly task-and finetuning data-dependent. We hope our findings could shed light on understanding, selecting and developing LLM finetuning methods. Advanced LLMs, such as GPT-4 (OpenAI, 2023) and PaLM 2 (Anil et al., 2023), often show emergent capabilities and allow for in-context learning that could use just a few demonstration examples to perform complex reasoning and generation tasks (Wei et al., 2022; Zhang et al., 2023; Fu et al., 2023; Shen et al., 2023). Still, LLM finetuning is required and widely adopted to unlock new and robust capabilities for creative tasks, get the most for focused downstream tasks, and align its value with human preferences (Ouyang et al., 2022; Yang et al., 2023; Gong et al., 2023; Schick et al., 2023). This becomes more significant in traditional industrial applications due to the existence of large-scale annotated task-specific data accumulated over years.

We conduct a large-scale fine-grained comparative analysis of machine translations (MT) against human translations (HT) through the lens of morphosyntactic divergence. Across three language pairs and two types of divergence defined as the structural difference between the source and the target, MT is consistently more conservative than HT, with less morphosyntactic diversity, more convergent patterns, and more one-to-one alignments. Through analysis on different decoding algorithms, we attribute this discrepancy to the use of beam search that biases MT towards more convergent patterns. This bias is most amplified when the convergent pattern appears around 50% of the time in training data. Lastly, we show that for a majority of morphosyntactic divergences, their presence in HT is correlated with decreased MT performance, presenting a greater challenge for MT systems.

Maximum-a-posteriori (MAP) decoding is the most widely used decoding strategy for neural machine translation (NMT) models. The underlying assumption is that model probability correlates well with human judgment, with better translations being more likely. However, research has shown that this assumption does not always hold, and decoding strategies which directly optimize a utility function, like Minimum Bayes Risk (MBR) or Quality-Aware decoding can significantly improve translation quality over standard MAP decoding. The main disadvantage of these methods is that they require an additional model to predict the utility, and additional steps during decoding, which makes the entire process computationally demanding. In this paper, we propose to make the NMT models themselves quality-aware by training them to estimate the quality of their own output. During decoding, we can use the model's own quality estimates to guide the generation process and produce the highest-quality translations possible. We demonstrate that the model can self-evaluate its own output during translation, eliminating the need for a separate quality estimation model. Moreover, we show that using this quality signal as a prompt during MAP decoding can significantly improve translation quality. When using the internal quality estimate to prune the hypothesis space during MBR decoding, we can not only further improve translation quality, but also reduce inference speed by two orders of magnitude.

We introduce PaLM 2, a new state-of-the-art language model that has better multilingual and reasoning capabilities and is more compute-efficient than its predecessor PaLM. PaLM 2 is a Transformer-based model trained using a mixture of objectives. Through extensive evaluations on English and multilingual language, and reasoning tasks, we demonstrate that PaLM 2 has significantly improved quality on downstream tasks across different model sizes, while simultaneously exhibiting faster and more efficient inference compared to PaLM. This improved efficiency enables broader deployment while also allowing the model to respond faster, for a more natural pace of interaction. PaLM 2 demonstrates robust reasoning capabilities exemplified by large improvements over PaLM on BIG-Bench and other reasoning tasks. PaLM 2 exhibits stable performance on a suite of responsible AI evaluations, and enables inference-time control over toxicity without additional overhead or impact on other capabilities. Overall, PaLM 2 achieves state-of-the-art performance across a diverse set of tasks and capabilities. When discussing the PaLM 2 family, it is important to distinguish between pre-trained models (of various sizes), fine-tuned variants of these models, and the user-facing products that use these models. In particular, user-facing products typically include additional pre- and post-processing steps. Additionally, the underlying models may evolve over time. Therefore, one should not expect the performance of user-facing products to exactly match the results reported in this report.