Towards Improving Translation Ability of Large Language Models on
Low Resource Languages
Amulya Ratna Dash and Yashvardhan Sharma
Birla Institute of Technology and Science, Pilani, Rajasthan, India
Keywords:
Natural Language Processing, Machine Translation, Low Resource Languages, Large Language Model.
Abstract:
With advancements in Natural Language Processing (NLP) and Large Language Models (LLMs), there is a
growing need to understand their capabilities with low resource languages. This study focuses on benchmark-
ing and improving the machine translation ability of LLMs for low resource Indic languages. We analyze the
impact of training dataset sizes and overfitting by training for additional epochs on translation quality. We
use LLaMA-3 as the base model and propose a simple resource efficient model finetuning approach which
improves the zero-shot translation performance consistently across eight translation directions.
1 INTRODUCTION
In recent years, Large Language Models are becom-
ing increasingly ubiquitous. LLM such as GPTs
(Brown et al., 2020), Gemma (Team et al., 2024),
LLaMA (Touvron et al., 2023) and others have shown
impressive performance in various NLP tasks like
text generation, question answering, summarization,
etc. Machine Translation (MT) helps break down lan-
guage barriers, provides access to global(foreign lan-
guage) information, and at times simplifies some day-
to-day tasks. As compared to other NLP tasks, ma-
chine translation involves the translation of text from
one language to another, which makes it relatively
complex. The accuracy of machine translation is crit-
ical as the user may not understand the foreign lan-
guage and completely rely on the generated text.
Most of the LLMs are pretrained on huge amount
of English language data along with very small
amount of non-English data (Minaee et al., 2024), and
this proportion is aligned with the available digital
text content. In comparison to supervised sequence-
to-sequence encoder-decoder Neural Machine Trans-
lation (NMT) models (Vaswani et al., 2017), the ma-
chine translation performance is lower for most of the
LLMs other than very large and closed language mod-
els like GPTs. The MT performance further degrades
for low resource languages. Given the growing pop-
ularity of LLMs, it is required to assess the transla-
tion ability of LLMs and explore ways to improve the
translation performance.
In this paper, we try to understand how well LLMs
perform machine translation for low resource Indic
languages and explore whether finetuning open LLMs
will improve the performance. Our findings show that
fine-tuning LLMs with limited parallel data and com-
putational resource can lead to significant improve-
ment in translation ability. We perform an ablation
study to evaluate how varying training dataset sizes
(low and medium) and the introduction of overfitting
through extended training epochs affect the model
performance.
2 RELATED WORK
2.1 LLMs for Machine Translation
Previous works have explored the use of LLMs for
MT, as these large models demonstrate language
understanding skills during pretraining (Liu et al.,
2024). In-context learning approach has been stud-
ied by Vilar et al., 2023, Agrawal et al., 2023 and
Bawden and Yvon, 2023, show the importance of few
shot examples, example quality and prompt template
for better MT performance. Hendy et al., 2023 have
performed a comprehensive evaluation of the transla-
tion quality of GPT models and found that the per-
formance for high resource languages is comparable
to dedicated translation systems. Singh et al., 2024
benchmarked LLMs on different generation tasks for
Indic languages, and reported PaLM-2 (a large closed
source LLM) leading in most tasks. Our work ex-
Dash, A. R. and Sharma, Y.
Towards Improving Translation Ability of Large Language Models on Low Resource Languages.
DOI: 10.5220/0013319000003905
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 801-807
ISBN: 978-989-758-730-6; ISSN: 2184-4313
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
801

plores zero shot performance of LLMs for MT of low
and very low resource Indic languages(and dialect).
2.2 Finetuning of LLMs
Limited prior work exists for adapting open LLMs
to machine translation tasks, with results similar to
closed models. Finetuning approach has been stud-
ied for relatively high resource languages and shows
that MT performance can be improved via finetuning.
Iyer et al., 2023 conclude that finetuning improves
the ability of LLMs to translate ambiguous sentences.
Xu et al., 2023 experiment on German, Czech, Rus-
sian, Chinese and Icelandic languages, adapt LLaMA
for translation using two steps, first with continued
pretraining on monolingual data and then finetuning
on parallel data. Alves et al., 2024 adapt LLaMa-
2 to multiple tasks in translation workflow, focus-
ing on German, French, Dutch, Italian, Spanish, Por-
tuguese, Korean, Russian and Chinese languages. In
our work, we focus on improving the MT perfor-
mance of LLaMA for low and very low resource In-
dic languages(and dialect) via single step finetuning
on limited parallel data presented as instructions.
3 METHODOLOGY
3.1 Models
To investigate the MT ability of LLMs, we use
three foundation language models: GPT 3.5 Turbo
1
,
Gemma 1.1
2
and LLaMA 3
3
. GPT 3.5 is a closed
model, where as Gemma and LLaMA are open mod-
els. Gemma 1.1 and LLaMA 3 offer two model sizes
each: 2 billion and 7 billion parameters for Gemma,
and 2 billion and 8 billion parameters for LLaMA. We
use the 7 billion model of Gemma and the 8 billion
model of LLaMA, henceforth referred to as Gemma
1.1 7B and LLaMA 3 8B.
3.2 Languages
To evaluate MT performance on low resource lan-
guages, we consider 4 Indic languages, i.e. Hindi,
Bengali, Odia and Chhattisgarhi. All four languages
belong to Indo-Aryan branch of the Indo-European
language family.
1
http://platform.openai.com/docs/models
2
https://huggingface.co/google/gemma-1.1-7b-it
3
https://huggingface.co/meta-llama/
Meta-Llama-3-8B-Instruct
Hindi is one of the main languages of India along
with English, and uses Devanagari script. Hindi is
spoken by 345 million speakers around the world.
Bengali (or Bangla) is the second most spoken
(237 million speakers) language in India and official
language of Bangladesh. It uses Bengali script and is
the fifth most spoken native language in the world.
Odia (or Oriya) language is spoken (50 million
speakers) predominantly in the Indian state of Odisha,
and uses Odia script.
Chhattisgarhi language is spoken (16 million
speakers) in the Indian state of Chhattisgarh and parts
of neighbouring states. It is considered as a dialect of
Hindi, at the same time due to its distinct linguistic
features, also considered as a separate language.
3.3 Evaluation Data
FLORES-200 (Costa-juss
`
a et al., 2022) is a human-
translated benchmark where the same English sen-
tences are translated into 200 languages. FLORES-
200 is split into dev(977), devtest(1012) and test(992)
categories. We use the publicly available devtest
split having 1012 sentences for evaluating MT per-
formance of LLMs, henceforth referred to as the test
dataset.
3.4 Evaluation Metrics
We use Character n-gram F-score(ChrF) (Popovi
´
c,
2015) and COMET-22 (Rei et al., 2022) scores as
evaluation metrics. ChrF evaluates translations at
character level rather than word or token level, mak-
ing it more robust for rich morphological and low
resource languages as compared to token-level met-
rics. Crosslingual Optimized Metric for Evaluation
of Translation 2022 (COMET 22) is a neural frame-
work to evaluate translations, which considers con-
text and semantics of the text and has better align-
ment with human judgement. COMET 22 is based
on the XLM-R model which doesn’t include Chhat-
tisgarhi language, so we evaluate Chhattisgarhi tasks
using ChrF only.
3.5 Finetuning of Model
Based on the baseline MT performance of Gemma
and LLaMA models on Hindi and Odia languages, we
select LLaMA as the base model for finetuning. We
finetune the LLaMA 3 8B model on parallel data pre-
sented as instructions using LoRA (Hu et al., 2021)
technique. LoRA is a technique that yields better per-
formance, improves sample efficiency, and reduces
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
802

Table 1: Composition of training dataset.
Language Pair Count(E2) Count(E3)
English - Hindi 1493 24846
Hindi - English 1504 24854
English - Bengali 1498 28724
Bengali - English 1499 28725
English - Odia 1489 18950
Odia - English 1508 18969
English - Chhattisgarhi 504 504
Chhattisgarhi - English
493 493
Total 9988 146065
reserved memory for finetuning LLMs to particular
tasks.
4 EXPERIMENTS
4.1 Tasks
We consider eight translation tasks grouped into two
categories. First category is Translation of test dataset
from English into Hindi, Bengali, Odia and Chhattis-
garhi, collectively referred to as en → xx. The second
category is translation of the test dataset from Hindi,
Bengali, Odia and Chhattisgarhi into English, collec-
tively referred to as xx → en.
4.2 Data
We collect our parallel training data from the Bharat
Parallel Corpus Collection (BPCC Gala et al., 2023)
and FLORES-200 dataset. BPCC-H Wiki and BPCC-
H Daily datasets which are manually translated multi-
domain Indic parallel corpus are used for Hindi,
Bengali, and Odia languages. The dev split of
FLORES-200 dataset is used for Hindi, Bengali, Odia
and Chhattisgarhi languages. The above individual
datasets are split into both translation directions for
uniformity across all 8 translation tasks. Data pro-
cessing includes formatting the data as per LLaMA
3 chat template. The composition of the processed
dataset is available in Table 1. The processed dataset
is split into train(0.9) and validation(0.1) dataset. The
template used for training data is available in Figure
1. A sample record from processed dataset is avail-
able in Figure 2.
4.3 Experiment Design
Experiment 1 (E1): Evaluate GPT3.5, Gemma1.1
7B and LLaMA3 8B models on the translation tasks.
Experiment 2 (E2): Finetune LLM on 10K parallel
data sampled from the processed dataset.
Table 2: Hyperparameters and Training size.
Parameters Values
Learning rate 0.0003
Train batch size 8
Eval batch size 4
Seed 3407
Gradient accumulation 8
Total train batch size 64
Optimizer Adam(Kingma,
2014)
Adam β
1
0.9
Adam β
2
0.999
Adam ε 1e-08
LR scheduler type cosine
LR warmup steps 0.01
Num of Epochs 4
Num of GPU 1 (A100 40GB)
Train examples
E2: 8989
E3: 131458
Experiment 3 (E3): Finetune LLM on the
full(100K+) dataset.
Both the experiments E2 and E3, include the step
to evaluate the finetuned model on the translation
tasks.
4.4 Training Setup
The model was finetuned in a multilingual many-to-
many approach, as is evident from the composition
of the training dataset. Among the open models, we
shortlisted LLaMA 3 8B for finetuning experiments
E2 and E3 based on the outcome of E1 experiment
(baseline). The hyperparameters and data size used
for experiments E2 and E3 are detailed in Table 2.
5 RESULTS
5.1 Baseline Experiment Results (E1)
In comparison to Gemma 1.1 model, GPT 3.5 and
LLaMa 3 perform better in the eight translation tasks.
Gemma 1.1 model performed poorly for the Odia
translation tasks. The scores for xx→en tasks are
better as compared to en→xx, as all three LLMs are
pretrained to generate fluent and coherent English
text. GPT 3.5 model performs better for Hindi tasks
and Hindi’s dialect Chhattisgarhi tasks. The detailed
scores from experiment E1 are available in Table 3.
Towards Improving Translation Ability of Large Language Models on Low Resource Languages
803

Figure 1: Template of training dataset.
Figure 2: Sample training record (and Prompt) for English→Odia language.
Table 3: Translation performance for GPT 3.5, Gemma 1.1, and LLaMA 3 (E1).
GPT 3.5 Gemma 1.1 LLaMA 3
Language Pair ChrF COMET ChrF COMET ChrF COMET
English - Hindi 47.80 77.78 40.08 74.51 43.77 76.57
English - Bengali 39.36 81.73 36.21 81.31 39.94 83.20
English - Odia 27.94 72.47 2.07 39.18 31.35 74.94
English - Chhattisgarhi 35.79 - 31.43 - 32.75 -
Hindi - English 60.37 87.94 55.95 86.21 60.10 87.86
Bengali - English 53.28 85.88 49.36 83.91 54.72 86.28
Odia - English 45.81 82.27 17.89 51.25 47.30 82.92
Chhattisgarhi - English 52.34 - 43.80 - 51.36 -
5.2 Finetuning Experiment Results (E2
and E3)
The objective of experiment E2 was to determine
the possibility of improving the translation perfor-
mance of LLaMa 3 by finetuning. Based on Valida-
tion loss, we select Epoch 2 checkpoint as the gen-
eralized model, and consider the same as the final
model of experiment E2. Using 10K parallel data and
1 hour of single GPU training time, we could see a
significant improvement in zero-shot translation per-
formance of the finetuned model. The average ChrF
score increased by +2.42 for xx→en and +6.14 for
en→xx over vanilla LLaMA 3 model. The detailed
scores from experiment E2 are available in Table 4
and Table 5.
As part of experiment E3, with around 12 hours
of single GPU training time and using 140K parallel
data spread across all 8 language pairs, we could see
further improvement in zero-shot translation perfor-
mance as compared to vanilla LLaMA 3 model. To
achieve the best generalization performance, we se-
Table 4: Translation performance on en → xx for model
trained with 10K examples (E2).
Language Pair ChrF COMET
English - Hindi 48.07 77.64
English - Bengali 43.74 83.34
English - Odia 38.21 77.22
English - Chhattisgarhi 42.36 -
Average 43.09 79.40
Table 5: Translation performance on xx → en for model
trained with 10K examples (E2).
Language Pair ChrF COMET
Hindi - English 59.71 87.98
Bengali - English 54.68 86.72
Odia - English 50.80 85.20
Chhattisgarhi - English 57.98 -
Average 55.79 86.63
lected the checkpoint from Epoch 2 based on the val-
idation loss. This checkpoint serves as our final fine-
tuned model. The average ChrF score increased by
+3.93 for xx→en and +11.37 for en→xx over vanilla
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
804

LLaMA 3 model. The detailed scores from experi-
ment E3 are available in Table 6 and Table 7.
Table 6: Translation performance on en → xx for model
trained with full dataset (E3).
Language Pair ChrF COMET
English - Hindi 51.94 78.91
English - Bengali 48.75 85.15
English - Odia 46.58 81.82
English - Chhattisgarhi 46.02 -
Average 48.32 81.96
Table 7: Translation performance on xx → en for model
trained with full dataset (E3).
Language Pair ChrF COMET
Hindi - English 61.06 88.03
Bengali - English 55.52 86.33
Odia - English 54.15 86.37
Chhattisgarhi - English 58.46 -
Average 57.30 86.91
6 ANALYSIS
As evident in Figure 3 and 4, the translation perfor-
mance of our finetuned model ‘IndicMT-Llama-3-8B’
is significantly better than 8B vanilla LLaMa 3 and
extremely large(more than 10x) closed model GPT
3.5 for both en→xx and xx→en translation directions.
Figure 3: Average zero-shot translation performance
(ChrF) on FLORES-200 for English ↔ {Hindi, Bengali,
Odia, Chhattisgarhi}.
6.1 Impact of Additional Parallel
Training Data
The final model trained with additional 130K paral-
lel training data had a roughly 2x improvement in
ChrF score. The improvement for en→xx was higher
as compared to xx→en. The additional training data
Figure 4: Average zero-shot translation performance
(COMET) on FLORES-200 for English ↔ {Hindi, Bengali,
Odia, Chhattisgarhi}.
helped the model better learn the Odia and Bengali
translation tasks, as seen by an improvement of up to
+8.37 ChrF score.
6.2 Impact of Increasing the Number of
Epochs
We analyzed the checkpoints of epoch 2 and
epoch 4 for the model finetuned with full training
dataset(140K). The ChrF score increased by +0.28 for
en→xx, but decreased by -5.37 for xx→en translation
direction. This mixed performance suggests that the
model overfits the training dataset when finetuned be-
yond epoch 2, forgetting its English text generation
ability.
6.3 Observations from Human
Evaluation
• Over-generation and repetition of words were ob-
served more when translating from English.
• Omission of words and generation of words from
the source (non-English) language appeared in
translated text in Roman script when translating
to English.
• The overall fluency of translated text was better in
the fine-tuned model compared to the base model
for both directions(to and from English).
• Table 8 shows few random translations from base
and finetuned model.
7 CONCLUSION AND FUTURE
WORK
As the field of NLP and LLMs progresses rapidly,
it is important to explore and improve their perfor-
Towards Improving Translation Ability of Large Language Models on Low Resource Languages
805

Table 8: Sample translation of FLORES-200 evaluation dataset using Base (B) and Finetuned (F) model.
English (en) Boating is a national pastime in Finland, with a boat to every seven or eight people.
B: Hindi → en Rowing is a national pastime in Finland, one boat for every seven or eight people.
F: Hindi → en Canoeing is a national pastime in Finland, with a boat accommodating one, seven or
eight people.
B: Bengali → en A boat ride with seven or eight people is a national pastime in Finland.
F: Bengali → en Boat races with seven or eight persons in a boat is a national pastime in Finland.
B: Odia → en It takes about 7 or 8 people to write a letter to Finland with a ferry.
F: Odia → en In Finland, canoeing is a national pastime, with a canoe for every 7 or 8 people.
B: Chhattisgarhi → en Finnland me donga me ghuma na ekthan rashtriya shagel ho, ek donga ke sang har
saat ya aath mankhem bar donga haway.
F: Chhattisgarhi → en In Finland, going by canoe is a national pastime, a canoe being the mode of transport
for one or two persons in a dongo.
mance on low resource languages. To address this
gap, this paper evaluates how well LLMs perform
machine translation for low resource Indic languages
and explores ways to improve the translation ability
of LLMs.
Our study concludes that open LLMs can be
finetuned to improve their multilingual translation
ability with very limited per language parallel data
and minimal computational resources. Our model
‘IndicMT-Llama-3-8B’ achieves superior translation
performance for four Indic languages, compared to
open LLMs, in both English-to-Indic and Indic-to-
English tasks. The methodology and findings can be
generalized to other languages in improving the trans-
lation performance of any LLM.
We have not included recent proprietary closed
models like GPT 4 (Achiam et al., 2023) and Gemini
(Team et al., 2023) in our study due to exceptionally
large model size as compared to less than 10B mod-
els like Gemma 1.1 7B & LLaMA 3 8B. The general
translation ability of our model beyond the four In-
dic languages should have improved, but we have not
validated.
Future work will evaluate the general translation
ability of our finetuned model, and extend the study to
additional languages(Indic and other language fami-
lies) and evaluation datasets. Research on methods to
further improve xx→en translation performance will
be helpful.
REFERENCES
Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I.,
Aleman, F. L., Almeida, D., Altenschmidt, J., Altman,
S., Anadkat, S., et al. (2023). Gpt-4 technical report.
arXiv preprint arXiv:2303.08774.
Agrawal, S., Zhou, C., Lewis, M., Zettlemoyer, L., and
Ghazvininejad, M. (2023). In-context examples se-
lection for machine translation. In Findings of the As-
sociation for Computational Linguistics: ACL 2023,
pages 8857–8873.
Alves, D. M., Pombal, J., Guerreiro, N. M., Martins, P. H.,
Alves, J., Farajian, A., Peters, B., Rei, R., Fernandes,
P., Agrawal, S., et al. (2024). Tower: An open mul-
tilingual large language model for translation-related
tasks. arXiv preprint arXiv:2402.17733.
Bawden, R. and Yvon, F. (2023). Investigating the trans-
lation performance of a large multilingual language
model: the case of bloom. In Proceedings of the 24th
Annual Conference of the European Association for
Machine Translation, pages 157–170.
Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,
Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,
Askell, A., et al. (2020). Language models are few-
shot learners. Advances in neural information pro-
cessing systems, 33:1877–1901.
Costa-juss
`
a, M. R., Cross, J., C¸ elebi, O., Elbayad, M.,
Heafield, K., Heffernan, K., Kalbassi, E., Lam, J.,
Licht, D., Maillard, J., et al. (2022). No language left
behind: Scaling human-centered machine translation.
arXiv preprint arXiv:2207.04672.
Gala, J., Chitale, P. A., AK, R., Gumma, V., Doddapa-
neni, S., Kumar, A., Nawale, J., Sujatha, A., Pudup-
pully, R., Raghavan, V., et al. (2023). Indictrans2: To-
wards high-quality and accessible machine translation
models for all 22 scheduled indian languages. arXiv
preprint arXiv:2305.16307.
Hendy, A., Abdelrehim, M., Sharaf, A., Raunak, V.,
Gabr, M., Matsushita, H., Kim, Y. J., Afify, M., and
Awadalla, H. H. (2023). How good are gpt models
at machine translation? a comprehensive evaluation.
arXiv preprint arXiv:2302.09210.
Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang,
S., Wang, L., and Chen, W. (2021). Lora: Low-rank
adaptation of large language models. arXiv preprint
arXiv:2106.09685.
Iyer, V., Chen, P., and Birch, A. (2023). Towards effec-
tive disambiguation for machine translation with large
language models. In Proceedings of the Eighth Con-
ference on Machine Translation, pages 482–495.
Kingma, D. P. (2014). Adam: A method for stochastic op-
timization. arXiv preprint arXiv:1412.6980.
Liu, Y., He, H., Han, T., Zhang, X., Liu, M., Tian, J., Zhang,
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
806

Y., Wang, J., Gao, X., Zhong, T., et al. (2024). Under-
standing llms: A comprehensive overview from train-
ing to inference. arXiv preprint arXiv:2401.02038.
Minaee, S., Mikolov, T., Nikzad, N., Chenaghlu, M.,
Socher, R., Amatriain, X., and Gao, J. (2024).
Large language models: A survey. arXiv preprint
arXiv:2402.06196.
Popovi
´
c, M. (2015). chrf: character n-gram f-score for auto-
matic mt evaluation. In Proceedings of the tenth work-
shop on statistical machine translation, pages 392–
395.
Rei, R., De Souza, J. G., Alves, D., Zerva, C., Farinha,
A. C., Glushkova, T., Lavie, A., Coheur, L., and Mar-
tins, A. F. (2022). Comet-22: Unbabel-ist 2022 sub-
mission for the metrics shared task. In Proceedings
of the Seventh Conference on Machine Translation
(WMT), pages 578–585.
Singh, H., Gupta, N., Bharadwaj, S., Tewari, D., and Taluk-
dar, P. (2024). IndicGenBench: A multilingual bench-
mark to evaluate generation capabilities of LLMs on
Indic languages. In Ku, L.-W., Martins, A., and
Srikumar, V., editors, Proceedings of the 62nd An-
nual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers), pages 11047–
11073, Bangkok, Thailand. Association for Computa-
tional Linguistics.
Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.-
B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A. M.,
Hauth, A., et al. (2023). Gemini: a family of
highly capable multimodal models. arXiv preprint
arXiv:2312.11805.
Team, G., Mesnard, T., Hardin, C., Dadashi, R., Bhupati-
raju, S., Pathak, S., Sifre, L., Rivi
`
ere, M., Kale, M. S.,
Love, J., Tafti, P., Hussenot, L., Sessa, P. G., Chowd-
hery, A., Roberts, A., Barua, A., Botev, A., Castro-
Ros, A., Slone, A., H
´
eliou, A., Tacchetti, A., Bu-
lanova, A., Paterson, A., Tsai, B., Shahriari, B., Lan,
C. L., Choquette-Choo, C. A., Crepy, C., Cer, D., Ip-
polito, D., Reid, D., Buchatskaya, E., Ni, E., Noland,
E., Yan, G., Tucker, G., Muraru, G.-C., Rozhdestven-
skiy, G., Michalewski, H., Tenney, I., Grishchenko, I.,
Austin, J., Keeling, J., Labanowski, J., Lespiau, J.-B.,
Stanway, J., Brennan, J., Chen, J., Ferret, J., Chiu, J.,
Mao-Jones, J., Lee, K., Yu, K., Millican, K., Sjoesund,
L. L., Lee, L., Dixon, L., Reid, M., Mikuła, M., Wirth,
M., Sharman, M., Chinaev, N., Thain, N., Bachem,
O., Chang, O., Wahltinez, O., Bailey, P., Michel, P.,
Yotov, P., Chaabouni, R., Comanescu, R., Jana, R.,
Anil, R., McIlroy, R., Liu, R., Mullins, R., Smith,
S. L., Borgeaud, S., Girgin, S., Douglas, S., Pandya,
S., Shakeri, S., De, S., Klimenko, T., Hennigan, T.,
Feinberg, V., Stokowiec, W., hui Chen, Y., Ahmed, Z.,
Gong, Z., Warkentin, T., Peran, L., Giang, M., Fara-
bet, C., Vinyals, O., Dean, J., Kavukcuoglu, K., Hass-
abis, D., Ghahramani, Z., Eck, D., Barral, J., Pereira,
F., Collins, E., Joulin, A., Fiedel, N., Senter, E., An-
dreev, A., and Kenealy, K. (2024). Gemma: Open
models based on gemini research and technology.
Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi,
A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava,
P., Bhosale, S., et al. (2023). Llama 2: Open foun-
dation and fine-tuned chat models. arXiv preprint
arXiv:2307.09288.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,
L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I.
(2017). Attention is all you need. Advances in neural
information processing systems, 30.
Vilar, D., Freitag, M., Cherry, C., Luo, J., Ratnakar, V.,
and Foster, G. (2023). Prompting palm for transla-
tion: Assessing strategies and performance. In Pro-
ceedings of the 61st Annual Meeting of the Associa-
tion for Computational Linguistics (Volume 1: Long
Papers), pages 15406–15427.
Xu, H., Kim, Y. J., Sharaf, A., and Awadalla, H. H. (2023).
A paradigm shift in machine translation: Boosting
translation performance of large language models.
arXiv preprint arXiv:2309.11674.
Towards Improving Translation Ability of Large Language Models on Low Resource Languages
807