Facilitating large language model Russian adaptation with Learned Embedding Propagation
Mikhail Tikhomirov, Daniil Chernyshev
TL;DR
The paper tackles the high cost of language-specific adaptation for open-source LLMs by introducing Learned Embedding Propagation (LEP), a lightweight embedding-level transfer method that avoids full instruction-tuning. LEP combines tokenization customization, embedding averaging initialization, and continued pre-training with optional post-training layer alignment to adapt models to Russian with minimal disruption. The authors also introduce the Darumeru benchmark to evaluate text generation robustness in offline, train-time settings and demonstrate that LEP, especially with vocabulary conversion projection, can match or exceed performance of instruction-tuned baselines on Russian tasks. The approach enables cost-efficient, open-source Russian LLM adaptation with practical benefits for sensitive environments and enables broader accessibility of language-specific capabilities.
Abstract
Rapid advancements of large language model (LLM) technologies led to the introduction of powerful open-source instruction-tuned LLMs that have the same text generation quality as the state-of-the-art counterparts such as GPT-4. While the emergence of such models accelerates the adoption of LLM technologies in sensitive-information environments the authors of such models don not disclose the training data necessary for replication of the results thus making the achievements model-exclusive. Since those open-source models are also multilingual this in turn reduces the benefits of training a language specific LLMs as improved inference computation efficiency becomes the only guaranteed advantage of such costly procedure. More cost-efficient options such as vocabulary extension and subsequent continued pre-training are also inhibited by the lack of access to high-quality instruction-tuning data since it is the major factor behind the resulting LLM task-solving capabilities. To address the limitations and cut the costs of the language adaptation pipeline we propose Learned Embedding Propagation (LEP). Unlike existing approaches our method has lower training data size requirements due to minimal impact on existing LLM knowledge which we reinforce using novel ad-hoc embedding propagation procedure that allows to skip the instruction-tuning step and instead implant the new language knowledge directly into any existing instruct-tuned variant. We evaluated four Russian vocabulary adaptations for LLaMa-3-8B and Mistral-7B, showing that LEP is competitive with traditional instruction-tuning methods, achieving performance comparable to OpenChat 3.5 and LLaMa-3-8B-Instruct, with further improvements via self-calibration and continued tuning enhancing task-solving capabilities.