Evaluating Large Language Models with Human

Feedback: Establishing a Swedish Benchmark

Birger Moëll
KTH
arXiv:2405.14006v1 [cs.CL] 22 May 2024

May 24, 2024

Abstract
In the rapidly evolving field of artificial intelligence, large language mod-
els (LLMs) have demonstrated significant capabilities across numerous appli-
cations. However, the performance of these models in languages with fewer
resources, such as Swedish, remains under-explored. This study introduces a
comprehensive human benchmark to assess the efficacy of prominent LLMs
in understanding and generating Swedish language texts using forced choice
ranking. We employ a modified version of the ChatbotArena benchmark,
incorporating human feedback to evaluate twelve different models, including
GPT-4, GPT-3.5, various Claude and Llama models, and bespoke models like
Dolphin-2.9-llama3b-8b-flashback and BeagleCatMunin. These models were
chosen based on their performance on LMSYS chatbot arena and the Scande-
val benchmarks. We release the chatbotarena.se benchmark as a tool to improve
our understanding of language model performance in Swedish with the hopes
that it will be widely used. We aim to create a leaderboard once sufficient data
has been collected and analysed.

1 Introduction
Large language models (LLMs) have demonstrated exceptional capabilities across
various applications, significantly advancing the field of natural language processing.
However, their effectiveness in low-resource languages remains underexplored. The
majority of these models are optimized for English or other high-resource languages,
leading to a notable performance disparity when applied to less commonly used lan-
guages. This gap in model performance has significant implications:

1
 • Accessibility: Individuals who are native speakers of low-resource languages
but not proficient in English are less able to benefit from the advancements in
AI and machine learning. This creates a barrier to accessing technology that
could otherwise support education, business, and communication.

• Research Limitations: The lack of scientific investigation into LLM perfor-

mance across diverse linguistic landscapes means that improvements and inno-
vations are often not tailored to the needs of all potential users. This oversight
can perpetuate inequities in technology access and effectiveness.

• Potential model improvements: Since models are not primarily tailored for
the language, improved understanding in combination with fine-tuning tech-
niques has the potential to significantly advance model performance for the
target language.

1.1 Motivation for creating Swedish Chatbot Arena

Benchmarks for Swedish exist with the excellent Scandeval [Nielsen(2023)] being
the primary tool used. ScandEval gives a good indication of model performance
but the reliance on automatic benchmarking has some drawbacks, training on the
benchmark can skew results and good performing models might not be what is pre-
ferred by humans. To adress these issues we introduce the Swedish Chatbot Arena,
a Swedish human preference benchmark, where the goal is to assess what models
humans prefer. Human can be seen as the gold standard for evaluating LLMs with
[Chiang et al.(2024)Chiang, Zheng, Sheng, Angelopoulos, Li, Li, Zhang, Zhu, Jordan, Gonzalez, and
LMSYS chatbot arena being the defacto standard for rating models reliably. Our
benchmark draws heavily on the LMSYS benchmark and we have adopted the source code
to work with our Swedish model.

1.2 Evaluation strategies

1.2.1 Subjective evaluation / Human feedback ranking
Human feedback plays a crucial role in the democratic evaluation of large language
models (LLMs). As these models increasingly influence various sectors such as media,
education, and even governance, it is essential to ensure that they operate in a manner
that aligns with the values and needs of diverse user groups. By involving humans in
the voting process on the quality of LLMs, we achieve several democratic objectives:

• Representation: Including a broad spectrum of individuals in the evaluation

process ensures that the models serve the needs of a diverse population rather

2
 than a select few. This inclusivity helps prevent biases that may otherwise
emerge if only a limited demographic is considered.

• Accountability: Human feedback mechanisms make developers and stake-

holders accountable to the end-users. By actively soliciting feedback, model
developers are encouraged to make improvements that are transparent and in
the best interest of the public.

• Adaptability: Through continuous human engagement, LLMs can be dynam-

ically adjusted to meet changing societal norms and expectations. This adapt-
ability is crucial in maintaining the relevance and appropriateness of automated
systems in a democratic society.

• Trust: When end-users contribute to the evaluation and enhancement of LLMs,

it builds trust in the technology. Trust is essential for the wider adoption and
acceptance of AI technologies in society.

• Ethical considerations: Human judgment is indispensable in navigating the

complex ethical landscapes that AI systems, like LLMs, often encounter. Vot-
ing on model quality allows for ethical deliberations that purely data-driven
approaches may overlook.

Swedish Chatbot Arena uses forced ranking between two models to create a bench-
mark of model performance using ELO ratings. This ensures a fair evaluations of
models while ensuring that a broad aspect of the public can be a part in democrati-
cally evaluating LLM performance.

1.3 Evaluation strategies

1.3.1 Objective evaluation / Automatic evaluation tools
Automatic evaluation tools such as Scandeval [Nielsen(2023)] play an indispensable
role in the development and refinement of large language models (LLMs). These
tools provide a systematic and standardized approach to assessing model performance
across multiple benchmarks, offering several distinct advantages:

• Efficiency: Automatic tools can evaluate models much more quickly than
human-based assessments. This speed is crucial for iterating over model designs,
allowing developers to refine and test models continuously without significant
delays.

3
 • Consistency: These tools apply the same standards and methodologies across
different models, ensuring that the evaluations are consistent. This uniformity
is essential for fair comparisons between models, facilitating a clearer under-
standing of each model’s strengths and weaknesses.

• Scalability: Automated tools can handle large volumes of evaluation tasks

simultaneously, making them scalable to the needs of rapidly developing AI
technologies. This capability is particularly valuable in the context of large-
scale LLMs, which require extensive testing across diverse datasets.

• Repeatability: Automatic evaluations can be repeated under the same condi-

tions to verify results. This repeatability helps in ensuring the reliability of the
evaluations and allows researchers to systematically track improvements over
time.

• Benchmarking: Objective evaluation provide benchmarks that are specifi-

cally designed to test various aspects of LLM performance, such as linguistic
understanding, contextual awareness, and the ability to generate coherent text.
This targeted assessment helps pinpoint specific areas where a model may need
further development.

The use of automated evaluation tools is crucial in the LLM development lifecycle
as they provide rapid feedback that is essential for timely and effective model training
and refinement. By leveraging these tools, developers can accelerate the development
process and enhance the overall quality and effectiveness of their models.

1.4 Objective and subjective evaluation

We believe that combining objective evaluation with a tool such as Scandeval with
subjective evaluation with a tool such as Swedish Chatbot Arena is the most efficient
way to get an accurate assessment of model performance.

2 Method
2.1 Evaluation platform
Our software evaluation platform is a fork of the LMSYS project which currently runs
on a RTX-4090 GPU with enough memory to serve two open source models.

4
2.2 Model Selection
Twelve language models were initially selected for inclusion based on evaluation on
LMSYS and Scandeval on Swedish-language tasks. Limitations such as availability in
the EU / availability through an API limited model selection. A majority of models
were served through an API, 10 models while two models Dolphin-2.9-llama3b-8b-
flashback and BeagleCatMunin was served through a local GPU.
The models currently included in the benchmark can be viewed in table 1.

Table 1: Language Models Selected

Model Model Size Open License LMSYS Scandeval Training
GPT-4o - No 1 - -
GPT-4-5 - No 2 - -
GPT-4 - No 2 1.07 -
GPT-3.5 - No 29 1.88 -
Claude Opus - No 2 - -
Claude Sonnet - No 8 - -
Claude Haiku - No 13 - -
Llama70b-Instruct 70b Yes 7 1.82 base + instruction
Llama-8b-Instruct 8b Yes 17 2.52 base + instruction
AI-Sweden-20b 20b No - 3.28 base + fine-tune
Dolphin-flashback* 8b Yes - 2.36 fine-tune + merge
BeagleCatMunin 8b Yes - 2.37 fine-tune + merge
*Full name, Dolphin-2.9-llama3b-8b-flashback
For Scandeval, lower rank is better.

2.3 Open AI models

The models GPT-3.5 and GPT-4, GPT-4.5 and GPT-4o was natural to include since
they are the most well used language models with GPT-3.5. powering ChatGPT and
GPT-4 being the most capable model in most evaluations and GPT-4o ranking num-
ber one on the LMSYS Chatbot Arena [Chiang et al.(2024)Chiang, Zheng, Sheng, Angelopoulos, Li, L

2.4 Anthropic Models

Anthropic released Anthropic version 3 on March 4th, 2024 with the models Haiki,
Sonnet and Opus. Training techniques includes pretraining on large data to acquire
language capabilities through word prediction, as well as human feedback techniques.

5
[Anthropic(2023)] Anthropic used a technique called Constitutional AI that uses
RLAIF for feedback for alignment. [Bai et al.(2022)Bai, Kadavath, Kundu, Askell, Kernion, Jones, C
The Opus overtook GPT-4 to become the best performing model on the LMSYS
benchmark before the release of GPT-4.5. The models are highly capable and al-
though they are not reviewed on Scandeval benchmark, they deserve to be reviewed
further in this evaluation.

2.5 Llama Models

On the 18th of April 2024 Meta open sourced Llama 3. Llama 3 was trained on 15
trillion tokens collected from public sources. Over 5% of the data consists of high
quality non English data. [Meta(2023)] Both the Llama 70b and the Llama 8b models
are capable with high scores on both LMSYS chatbot arena and Scandeval.

2.5.1 Dolphin-2.9-llama3b-8b-flashback
Several fine-tunes and merges have been done on Llama3 models for Swedish. Dolphin-
2.9-llama3b-8b-flashback is a high performing example of a lllama-3 fine tune. The
model is a merge of timpal0l/Llama-3-8B-flashback-v1 and cognitivecomputations/dolphin-
2.9-llama3-8b which in turn are fine-tunes on Swedish language data (Flashback) and
English language data. Notably the underlying dolphin model is uncensored to re-
move alignment and bias. This model is part of the the benchmark for research
purpose but can generate unethical responses.

2.6 Open models

2.6.1 BeagleCatMunin
BeagleCatMunin is similar to Dolphin-2.9-llama3b-8b-flashback by being a merge of
fine-tuned models. The model is based on the Mistral-7b architecture. The underlying
Swedish model has been trained on flashback data and it is merged with a highly
performant Scandinvian model, RJuro/munin-neuralbeagle-7b.

2.6.2 AI-Sweden-20B
AI-Sweden 20b is a model created by AI Sweden and trained on the Nordic Pile
dataset.

6
3 Types of models evaluated
Since the benchmark contains both open source and closed source, not all knowledge
regarding training of the models is available. However, based on information regarding
the open models some conclusions regarding models can be drawn.

3.1 Instruction tuned models

Instruction tuning involves creating a dataset of instructions (prompts) and answer
for a base trained model to follow. Instruction tuning has been shown to increase
model performance and Llama 8b Instruct and Llama 70b Instruct are examples of
instruction tuned models in this benchmark.

3.2 Reinforcement Learning from Human Feedback

Reinforcement Learning from Human Feedback is a step in training where a model
uses human feedback to improve model performance. Most models included in this
benchmark has at some point been trained with RLHF. Notably Dolphin-2.9-llama3b-
8b-flashback is the only model where an attempt has been made to remove preferences
made from RLFH training.

3.3 Merged models

Merging[Yadav et al.(2023)Yadav, Tam, Choshen, Raffel, and Bansal] is a technique
where models are combined to form a new model by combining the weighs inside
the neural network directly. This data free training method has several benefits with
perhaps the biggest one being the ability to merge a model without the need for a
GPU. This in combination with the high quality performance of merge models have
made them a common technique for working with models.

4 Next steps
We hope that everyone will help out in our work on improving Swedish Language
models by evaluating models on chatbotarena.se, the home of our benchmark.
Once evaluations are done and we have gathered enough data, we will present a
benchmark of the most performant models for Swedish.

7
4.1 Data
Our benchmark relies on forced binary ranking of two models which gives as output
a preference for an answer over another. This output is useful both as assessment of
model performance but also as a way to create Human Preference Data that can be
used to improve model performance.

4.2 How can you help?

The goal of our project is to improve access to high quality language models in
Swedish. The easiest way to help is to simply evaluate models. We also welcome pull
requests and feedback on our repo1 .

References
[Anthropic(2023)] Anthropic. Model card for claude 3.
https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card
2023. Accessed: 2024-05-02.

[Bai et al.(2022)Bai, Kadavath, Kundu, Askell, Kernion, Jones, Chen, Goldie, Mirhoseini, McKinnon
Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson
Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron
McKinnon, Carol Chen, Catherine Olsson, Christopher Olah, Danny Hernandez,
Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-Johnson, Ethan Perez, Jamie
Kerr, Jared Mueller, Jeffrey Ladish, Joshua Landau, Kamal Ndousse, Kamile
Lukosuite, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noemi
Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott
Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham,
Timothy Telleen-Lawton, Tom Conerly, Tom Henighan, Tristan Hume, Samuel R.
Bowman, Zac Hatfield-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, Sam
McCandlish, Tom Brown, and Jared Kaplan. Constitutional ai: Harmlessness
from ai feedback, 2022.

[Chiang et al.(2024)Chiang, Zheng, Sheng, Angelopoulos, Li, Li, Zhang, Zhu, Jordan, Gonzalez, and
Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos,
Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael Jordan, Joseph E.
Gonzalez, and Ion Stoica. Chatbot arena: An open platform for evaluating llms
by human preference, 2024.
1
https://github.com/BirgerMoell/SwedishLLMBenchmark

8
[Meta(2023)] Meta. Llama 3 release blog. https://ai.meta.com/blog/meta-llama-3/,
2023. Accessed: 2024-05-02.

[Nielsen(2023)] Dan Nielsen. ScandEval: A benchmark for Scandinavian natural

language processing. In Tanel Alumäe and Mark Fishel, editors, Proceedings of
the 24th Nordic Conference on Computational Linguistics (NoDaLiDa), pages
185–201, Tórshavn, Faroe Islands, May 2023. University of Tartu Library. URL
https://aclanthology.org/2023.nodalida-1.20.

[Yadav et al.(2023)Yadav, Tam, Choshen, Raffel, and Bansal] Prateek Yadav, Derek
Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. Ties-merging: Resolving
interference when merging models, 2023.