Although still in their infancy, large language models (LLMs) have emerged as a tool with the potential to transform human-computer interaction and significantly impact various aspects of work and daily life (see e.g., Bubeck et al. (2023), etc.).
Due to this widespread use and the existence of a wide variety of language models, it is crucial to establish a standardized method for evaluating and ranking these models based on their performance. Improved evaluation will provide guidance for future interaction design and implementation. There are multiple general approaches for evaluating evaluate the performance of LLMs.
The first approach is a static approach which involves evaluating models based on a fixed set of pre-determined questions (benchmarks). Many such benchmarks have been proposed to evaluate the performance of LLMs in various domains, such as medical applications (Singhal et al. (2023); Cas-cella et al. (2023); Liévin et al. ( 2024)), legal applications (Hendrycks et al. (2021); Guha et al. (2024); Katz et al. ( 2024)), trustworthiness (Chao et al. (2024); Zhang et al. (2023)), reasoning abilities (Sawada et al. (2023); Valmeekam et al. (2022)), and coding abilities (Liu et al. (2024a); Du et al. (2024); Carlini (2024)). See Chang et al. (2024) for a detailed survey. However, these benchmarks are limited, mainly because they may become contaminated over time as new language models are introduced with the benchmarks potentially included as their training data (see e.g., Bubeck et al. (2023); Ibrahim et al. (2024)). Thus, rankings derived using these methods may not generalize to other, new tasks and questions, even within the same domain.
The second approach to LLM evaluation is a human-based approach in which human evaluators are asked to interact with and compare models by prompting, then ranking their performance based on their responses. An example is Chatbot Arena, recently introduced by Chiang et al. (2024), which evaluates LLMs using human feedback collected through crowd-sourcing. Chatbot Arena has attracted significant attention and media coverage (see, e.g., Yang and Cui (2024); Roose (2024)). Human-based approaches can effectively resolve the problem of data contamination in the static approaches. However, their reliance on human input limits their scalability. Designing suitable prompts and reading (often long) responses from various models can be very expensive and time-consuming.
The third approach to model evaluation is a game-based approach which bypasses the need for human evaluation by designing a structured game in which models compete against each other. The game is crafted with automatically checkable winning criteria, so the winner can be determined automatically without human intervention. The models are then ranked based on their performance in the game. The game should be designed in a way to require the relevant skills expected of language mod-els so that the performance in these games serves as a proxy for the model's overall abilities. As a result, a good choice of the game is very important for the success of game-based approaches.
Debate has been a longstanding tradition since antiquity (Dutilh Novaes, 2022, Historical Supplement). It serves as a structured forum for testing conversational and reasoning skills, and has been integral to philosophical discourse, legal proceedings, and civic engagement (Proksch and Slapin (2015); Holbrook (1999); Benoit et al. (2003)). Although success in debates requires mastery of the debate topic and domain knowledge, it also needs skills such as defining the problem, identifying and challenging assumptions, recognizing inconsistencies, and prioritizing the relevance of various details within the overall argument (Roy and Macchiette, 2005;Kennedy, 2009). Such skills are crucial for the effective application of LLMs, which motivates using debate settings in game-based approach to LLM evaluation.
In this paper, we take the game-based approach to model evaluation and design the game to be a structured debate between competing language models, judged by a pre-specified language model. In this framework, models debate on a set of predetermined topics. The script of each debate is then given to a pre-determined judge, which is also an LLM, to evaluate and score the arguments presented by each side. Based on these evaluations, the language model with the better overall performance is announced as the winner. Through experiments with state-of-the-art language models, we demonstrate that this fully automated setup can be used to rank these LLMs, producing rankings generally consistent with those of Chatbot Arena, which is one of the most widely used methods of LLM evaluation.
Collaboration and debates have been used as a framework to enhance the performance of language models. Liang et al. (2023); De Ridder (2024) propose multi-agent debate frameworks in which models express their solutions to a problem to encourage divergent thinking. Xiong et al. (2023) uses a formal debate framework to examine whether LLMs can collaborate to reach a consensus for a shared goal. Chern et al. (2024) proposes multiround discussions between models to assist human Figure 1: A snippet of debates. Two language models engage in debates on a list of topics, and a judge model announces the winner for each topic. The language model with the most wins across all topics is declared the overall winner.
annotators in finding the most capable LLM to be used as an evaluator. Khan et al. (2024) shows that having access to the a debate between two strong LLMs that posses the necessary information to answer the question of the topic will result in more truthful answers from a weak LLM or a human expert that do not have that piece of information. Lee et al. (2024) (2024) for other results leveraging the interactions between language models to achieve a given goal.
Assume that T is a pre-defined list of debate topics and LM is a set of large language models that we want to rank based on their performance.
The topics are open-ended questions like "Can alternative energy effectively replace fossil fuels?" that have two possible sides. We assume that we have black-box query access to all the language models. The ranking is based on multiple debates between the language models on different topics.
Each debate is a multi-round interaction on a topic t ∈ T between two models, a language model LM 1 ∈ LM that goes first and support one side of the argument, and a language model LM 2 ∈ LM that goes second and supports the other side. In the first round, LM 1 is asked to start the debate by providing the arguments to supports their side based on logic, facts, and evidence. Then, in the next round LM 2 responds, refutes the arguments raised, and provides new evidence supporting the other side. The debate between the models continues for a predefined number T of rounds. Finally, the model LM 2 is asked to conclude the debate. All prompts used in our experiments can be found in Section A. At the end of each debate, the script of the T rounds is given to a judge LLM, which is asked to consider specific pre-defined factors, to score LM 1 and LM 2 , and to announce the winner.
In this framework, to compare two models LM a and LM b from the set L, we conduct two debates on each topic t ∈ T . In the first debate, we set LM 1 = LM a , and in the second debate, we set LM 1 = LM b . The judge language models evaluate both debates. Based on their assessments, one of the models LM 1 or LM 2 is declared the winner for the topic t, or the result is a draw. We run two debates because the two sides of the argument might not be equally hard to argue for. Also, the judge could be biased and favor the model that goes first (or last). By running the debate with the role of the models flipped, we account for these biases. This process is repeated for all topics t ∈ T . The model with the highest number of wins across different topics is declared the overall winner. Finally, after comparing all pairs (LM a , LM b ) of models, the overall ranking of the models in L is generated.
Since LLMs are typically not directly trained to debate, specifying the rules of the task is crucial. To effectively do this, we provide a detailed system prompt to the models. The systems prompts ask the models to support a side; ask that the arguments and rebuttals should be backed by logic, facts, and evidence; and that the answers should be convincing, factual and concise. The history of all previous rounds of the debate is given to the debating LLMs.
The same is also true for the judge language model. We set a detailed system prompt for the judge so that it considers factors such as clarity of arguments, factuality and use of evidence, rebuttal and counterarguments, logical consistency, persuasiveness, conciseness, and coherence in the evaluation. Further, in the system prompt for the judge, we specify the exact format for the output. The script of the debate is given to the judge model using its prompt and the judge model announces a winner.
LLMs have been used as judges for various applications (see e.g., Zheng et al. (2024);Chen et al. (2024);Chao et al. (2024); Kim et al.;Hada et al. (2024b,a); Wu and Aji (2023); Kim et al. (2024)).
Because of the symmetry in the game, the evaluation based on the game is fair regardless of the strength or weakness of the judge model. However, when we use weaker language models as a judge, we give preference to language models that are stronger at convincing the judge that their argument is coherent, instead of the models that are actually giving coherent responses. In other words, models that are able to generate responses that are marked by the judge to be more coherent are scored higher. Note that the same is true for human judges. For example, in Chatbot Arena, users select the response of an LLM not necessarily based on whether it is more "coherent", but based on whether it is more "coherent". Although this might not be the right metric for all applications, we argue that for the conversation-based tasks that chatbots are used for everyday, the persuasiveness is the key ability users seek when choosing the language model. For example, see OpenAI o1 System Card (OpenAI, 2024, Section 4.7.1).
For the task of evaluating debates, Liu et al. (2024b) showed that of GPT-4 outperforms humans and other state-of-the-art LLMs fine-tuned on extensive datasets in debate evaluation. More generally, GPT-4 models have consistently been demonstrated to closely match with human intentions when acting as a judge and have been used as a judge extensively in LLM literature (see e.g., Zheng et al. (2024); Achiam et al. (2023) and references therein). Based on these findings, we choose GPT-4 as the judge model in our experiments. In Section 4.2, conduct some experiments with Llama-3-70b as the judge and demonstrate that the rankings do not change significantly when Llama-3-70b is used as judge. Also, we conducted an experiment where a human judge also asked to determine the winner of debates by reading them. We showed that the winners chosen by GPT-4 is consistent with the winners chosen by the human evaluator.
In this paper, we evaluate the overall conversation ability of language models with general questions and use general-purpose LLMs as judge. However, we note that if all the debate topics are from a specific subject, choosing a Retrieval-Augmented Generation LLM (see e.g., (Lewis et al., 2020)) with a proper knowledge-base might be better suited for determining the factuality of the claims by different sides of the debate.
In this section, we consider the list of debate topics from Section B and run debates with four rounds, i.e., T = 4. As the debating models, we use Llama-2-7b, Llama-2-13b, Llama-2-70b (Touvron et al., 2023), Llama-3-70b (Meta AI, 2024), Vicuna-7b-v1.5,Vicuna-13b-v1.5 (Chiang et al., 2023), Mixtral-8x7B-Instruct-v0.1 (Jiang et al., 2024), gpt-4-0125-preview, and gpt-3.5-turbo-0125 (Achiam et al., 2023). To access GPT models, we use the OpenAI API. For other models, we use the API from https://www.together.ai/.
We run a total of 50 debates on 25 topics (Section B) between each pair of models. In this section we use gpt-4-0125-preview as the judge. On each topic, the model that wins both rounds is considered the winner. Table 1 (Left) shows the number of wins in the debates between various model pairs. For example, Llama-2-7b has won in no topics against Llama-2-70b, whereas Llama-2-70b has won in nine debates against Llama-2-7b. The detailed results in different topics can be found in Section C. See Table 1 and Figure 2 From these results, as a sanity check, it is seen that for different families of models (Llama-2, GPT, Vicuna), models with more parameters rank better; i.e., Llama-2-70b ranks better compared to Llama-2-7b. Also, newer generations of each model rank better than their older counterpart; i.e., GPT-4 ranks better compared to GPT-3.5. Similarly, Llama-3-70b ranks better compared to Llama-2-70b.
Comparing this ranking with the rankings available on the Chatbot Arena leaderboard website, accessed on June 14th, 2024, we see that the rankings in Table 1 are generally consistent with Chatbot Arena. Specifically, the normalized Kendall tau distance between these two rankings is 0.0833. This distance takes values in [0, 1] where 0 means identical rankings and 1 means reversed rankings. Llama-2-7b 1-0 0-6 0-9 0-11 9-0 6-2 1-8 0-5 0-24 Llama-2-13b 0-2 0-7 2-13 14-0 8-0 0-5 0-7 0-22 Llama-2-70b
1-1 0-9 13-0 12-0 2-1 1-2 0-21 Llama-3-70b 0-1 23-0 17-0 6-0 5-0 0-13 Vicuna-7b 2-1 0-3 0-11 0-18 0-24 Vicuna-13b 0-0 0-13 0-13 0-23
Analysis of Content. We prompt the judge (GPT-4) model to state the main reason for its choice among "clarity, factuality, counterarguments, persuasiveness, conciseness, and coherence." Figure 3 illustrates the stated reasons for which different models won the debates against their opponents. This shows that clarity and coherence have been the most decisive factors in the decisions. It also reveals that, for example, the responses by GPT-4 were seen by the judge as being more coherent, whereas the arguments by Vicuna-7b were seen to be more concise. 0% 25% 50% 75% 100% L la m a 2 -7 b L la m a 2 -7 0 b L la m a 3 -7 0 b V ic u n a -7 b V ic u n a -1 3 b G P T 3 .5 G P T 4 Coherence Conciseness Persuasiveness Counterarguments Factuality Clarity Human as Judge. In this experiment, we asked three human student volunteers to read the contents of the debates between Llama-2-13b and Llama-2-70b and judge the debates. The human evaluators each judged a total of 50 debates and their judgments matched the results of Section C.11 in 81.3% of the time. In particular, each participant agreed with the judgment from Section C.11 in 43, 41, and 38 debates out of 50.
Llama-3 as Judge. Finally, we conduct similar experiments, but with Llama-3-70b as the judge.
For demonstration, we only repeat the debates between Llama-2-13b with all other models. The score of Llama-2-13b against other models with Llama-3-70b as judge is shown in Table 2. The winners in this experiment are identical to the winners announced by GPT-4, in all but one opponent models (Mixtral).
Table 2: Score of Llama-2-13b vs. other models (first numbers for Llama-2-13b), with Llama-3-70b as judge. Model Score Model Score Llama-2-7b 7-0 Vicuna-13b 17-0 Llama-2-70b 0-4 Mixtral-8x7B 8-4 Llama-3-70b 0-6 GPT-3.5 0-4 Vicuna-7b 16-2 GPT-4 0-21
In this paper, we developed an automated framework to rank the performance of LLMs, based on a multi-round debate between LLMs on different topics, and an evaluation by a judge LLM. We showed that this framework can yield rankings consistent with rankings that rely on human crowdsourcing, while being more scalable.
List of Topics. The method proposed in this paper requires human input to create the list of debate topics. While this is significantly less timeconsuming and less expensive than reading and scoring the debates, it can still pose some scalability issues. We leave the task of automating topic generation for future work.
LLM as Judge. The debate framework introduced in this paper is a game-based approach to model evaluation and uses a judge LLM as a part of the game. Because of the symmetry in the game design, the judge cannot bias the game; however, if a weak LLM is chosen as the judge, it is possible that the judge may not be capable of fully evaluating qualities such as the "factuality" of the arguments. In such cases, the winner is determined by the abilities of different models to convince the judge that they are more factual, instead of actually giving more factual answers. Although this still demonstrates the abilities of that language model, it might not necessarily be the qualities that we try to evaluate. Note that the same is true for human judges. For example, in Chatbot Arena, users select an LLM not necessarily based on whether it is generated more "factual" responses, but based on whether it can generate more "convincing" outputs.
In our experiments in Section 4.2, it was shown that the decisions are mostly based on factors such as coherence, conciseness, or clarity that are easier for a language model to evaluate. However, despite the evidence (see e.g., Liu et al. (2024b)) that models such as GPT-4 perform well on debate evaluation tasks, in general this limits the applicability of the debate framework to evaluate the ability of models in tasks other than the general conversational ability that was studied here, where factors such as factuality are more important.
Choice of Language. Our method has been primarily evaluated on English, which is a language with relatively limited morphological complexity. This could restrict the applicability of our approach to other languages.
In this section, we list the system prompts and prompts used in the experiments.
The first language model LM 1 is asked to start the debate. This language model will have the following system prompt.
We are having a debate and the topic is "TOPIC". You are representing "SIDE 1" and you should zealously support it. This the first round and you are going first. You must bring-up arguments supporting your side backed by logic, facts and evidence. Your answer should also be convincing, factual and concise.
We will prompt LM 1 as follows.
Please start the debate.
In the second part of the debate, the second language model LM 2 will be given the response of LM 1 . The second model is asked to provide a rebuttal to the points raised by LM 1 . Also, it is asked to provide new arguments supporting the second side. This language model will have the following system prompt.
We are having a debate and the topic is "TOPIC". You are representing "SIDE 2" and you should zealously support it. The other side has started the debate and you will be given their arguments. You must first provide a rebuttal to the points raised by them, and then provide new arguments supporting your side. All your arguments should be backed by logic, facts and evidence. Your answer should also be convincing, factual and concise.
We prompt LM 2 as follows.
The other side said: "Response of LM 1 in Part 1".
In all following steps, we use the following system prompts for the models.
We are having a debate and the topic is "TOPIC". You are representing "SIDE" and you should zealously support it. You will be given all the arguments so far. In your response, you should support your side and refute the points raised by the other side. Your arguments should be backed by logic, facts and evidence. Your answer should also be convincing, factual and concise.
We prompt the models LM 1 and LM 2 as follows.
You initially said: "Response of LM 1 in Part 1". You the other side responded: "Response of LM 2 in Part 2". Then you said: "Response of LM 1 in Part 3", . . . .
The other side initially said: "Response of LM 1 in Part 1". You then responded: "Response of LM 2 in Part 2". The other side said: "Response of LM 1 in Part 3", . . . .
A judge language model gives scores from 1 to 10 to each debate participant based on the following criteria: (1) Clarity of arguments (2) Factuality and use of evidence (3) Rebuttal and counterarguments (4) Logical consistency (5) Persuasiveness and impact (6) Conciseness (7) Coherence. The judges will have the following system prompt.
We had a debate and the topic was "TOPIC". The two sides in the debate each provided arguments to prove their side and refute the points raised by the opponent. You are a judge for this debate. You should be impartial and as objective as possible. The debate script will be given. You should give a score from 1 to 10 to each side of the debate. In your judgement, you should take into account the following criteria: clarity of arguments, factuality and use of evidence, rebuttal and counterarguments, logical consistency, persuasiveness and impact, conciseness, coherence. Also, you should choose the side who you think is the overall winner. Your answer MUST follow the following format: "side1: [[score of side 1]], side2: [[score of side 2]], winner: [[name of winner]]" where score of side 1 and score of side 2 are numbers from 1 to 10 and name of winner is either "1" or "2".
The judge generates the scores given the whole script of the debate, using the following prompt.
The script of the debate is as follows: Side 1: "Response of LM 1 in Part 1". Side 2: "Response of LM 2 in Part 2". Side 1: "Response of LM 1 in Part 3". Side 2: "Response of LM 2 in Part 4".
The twenty-five debate topics used in the experiments are as follows.
1. Can alternative energy effectively replace fossil fuels? 2. Should K-12 students dissect animals in science classrooms? 3. Is artificial intelligence good for society? 4. Should bottled water be banned? 5. Is a college education worth it? 6. Should the United States keep daylight saving time? 7. Should school dress codes be implemented and enforced? 8. Should the drinking age be lowered from 21 to a younger age? 9. Should the election day be made a national holiday? 10. Should the governments use Large Language Models for advice? 11. Should employers be able to mandate vaccinations? 12. Should fighting be allowed in hockey? 13. Should fur clothing be banned? 14. Should genetically modified organisms (GMOs) be grown? 15. Is golf a sport and are golfers athletes? 16. Is homework beneficial? 17. Is the internet "making us stupid?" 18. Should medical aid in dying be legal? 19. Is obesity a disease? 20. Should the penny stay in circulation? 21. Are the Olympic games an overall benefit for their host countries and cities? 22. Is there really a Santa Claus? 23. Should Halloween be moved permanently to Saturday? 24. Should students have to wear school uniforms? 25. Is social media good for society?
In this section, we report the experimental results with GPT-4 as judge. Each table is the result of the debates between two models on all topics. Home and Away correspond to to runs of the debate on each topic, each time with one LLM going first.
Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-7b Llama-2-7b Winner Llama-2-7b Llama-2-7b 1 Side 1 8 7 Side 1 9 2 Side 2 7 8 Side 2 8 3 Side 1 8 7 Side 1 8 4 Side 1 8 7 Side 1 8 5 Side 1 8 7 Side 2 8 6 Side 1 8 7 Side 1 8 7 Side 2 7 8 Side 2 7 8 Side 2 7 8 Side 2 6 9 Side 1 8 7 Side 1 8 10 Side 1 8 7 Side 1 8 11 Side 1 9 7 Side 1 8 12 Side 2 7 8 Side 2 7 13 Side 1 8 7 Side 1 8 14 Side 1 8 7 Side 1 8 15 Side 1 8 6 Side 1 8 16 Side 2 7 8 Side 2 7 17 Side 2 7 8 Side 2 7 18 Side 1 8 7 Side 1 8 19 Side 1 9 7 Side 1 9 20 Side 2 7 8 Side 2 7 21 Side 2 7 8 Side 2 8 22 Side 2 6 8 Side 2 6 23 Side 1 8 7 Side 1 8 24 Side 2 7 8 Side 2 7 25 Side 1 8 7 Side 1 8 2049 C.2 Llama-2-7b vs Llama-2-13b Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-13b Llama-2-7b Winner Llama-2-7b Llama-2-13b Overall 1 Llama-2-13b 8 7 Llama-2-7b 8 7 Tie 2 Llama-2-7b 7 8 Llama-2-13b 7 8 Tie 3 Llama-2-13b 8 7 Llama-2-7b 8 7 Tie 4 Llama-2-13b 8 7 Llama-2-7b 8 7 Tie 5 Llama-2-13b 8 7 Llama-2-7b 8 7 Tie 6 Llama-2-13b 8 7 Tie 8 8 Llama-2-13b 7 Llama-2-7b 8 9 Llama-2-13b 7 9 Tie 8 Llama-2-7b 6 8 Llama-2-13b 7 8 Tie 9 Llama-2-13b 8 7 Llama-2-7b 8 7 Tie 10 Llama-2-13b 8 7 Llama-2-7b 8 7 Tie 11 Llama-2-13b 8 7 Llama-2-7b 8 7 Tie 12 Llama-2-7b 7 8 Llama-2-13b 7 8 Tie 13 Llama-2-13b 8 7 Llama-2-7b 8 7 Tie 14 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 15 Llama-2-13b 8 7 Llama-2-7b 8 7 Tie 16 Llama-2-13b 8 7 Llama-2-7b 8 7 Tie 17 Llama-2-7b 6 8 Llama-2-13b 7 8 Tie 18 Llama-2-13b 8 7 Llama-2-7b 8 7 Tie 19 Llama-2-13b 9 7 Llama-2-7b 8 7 Tie 20 Llama-2-7b 6 8 Llama-2-13b 7 8 Tie 21 Llama-2-13b 8 7 Llama-2-13b 8 9 Llama-2-13b 22 Llama-2-7b 6 8 Llama-2-13b 6 8 Tie 23 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 24 Llama-2-13b 8 7 Llama-2-13b 8 9 Llama-2-13b 25 Llama-2-13b 8 7 Llama-2-13b 8 9 Llama-2-13b C.3 Llama-2-7b vs. Llama-2-70b Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-7b Llama-2-70b Winner Llama-2-70b Llama-2-7b Overall 1 Llama-2-7b 8 7 Llama-2-70b 9 Tie 2 Llama-2-70b 7 9 Llama-2-7b 7 Tie 3 Llama-2-7b 8 7 Llama-2-70b 8 Tie 4 Llama-2-7b 8 7 Llama-2-70b 8 Tie 5 Llama-2-70b 8 9 Llama-2-70b 8 Llama-2-70b 6 Llama-2-70b 7 8 Llama-2-70b 8 Llama-2-70b 7 Llama-2-70b 7 8 Llama-2-70b 8 Llama-2-70b 8 Llama-2-70b 6 8 Llama-2-7b 7 Tie 9 Llama-2-7b 8 7 Llama-2-70b 8 Tie 10 Llama-2-70b 8 8 Llama-2-70b 8 Llama-2-70b 11 Llama-2-7b 8 7 Llama-2-70b 8 Tie 12 Llama-2-70b 7 9 Llama-2-7b 8 Tie 13 Llama-2-70b 7 8 Llama-2-70b 8 Llama-2-70b 14 Llama-2-7b 8 7 Llama-2-70b 8 Tie 15 Llama-2-7b 8 7 Llama-2-70b 8 Tie 16 Llama-2-7b 8 8 Llama-2-70b 8 Tie 17 Llama-2-70b 7 8 Llama-2-7b 7 Tie 18 Llama-2-7b 8 7 Llama-2-70b 8 Tie 19 Llama-2-7b 9 8 Llama-2-70b 8 Tie 20 Llama-2-70b 6 8 Llama-2-7b 7 Tie 21 Llama-2-70b 8 9 Llama-2-70b 8 Llama-2-70b 22 Llama-2-70b 6 8 Llama-2-70b 8 Llama-2-70b 23 Llama-2-7b 8 7 Llama-2-70b 8 Tie 24 Llama-2-70b 7 8 Llama-2-70b 8 Llama-2-70b 25 Llama-2-70b 7 8 Llama-2-70b 9 Llama-2-70b 2050 C.4 Llama-2-7b vs. Llama-3-70b Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-3-70b Llama-2-7b Winner Llama-2-7b Llama-3-70b Overall 1 Llama-3-70b 9 8 Llama-2-7b 9 8 Tie 2 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 3 Llama-2-7b 8 9 Llama-3-70b 8 9 Tie 4 Llama-3-70b 8 6 Llama-2-7b 8 7 Tie 5 Llama-3-70b 8 7 Llama-2-7b 8 7 Tie 6 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 7 Llama-3-70b 8 7 Llama-3-70b 7 9 Llama-3-70b 8 Llama-3-70b 8 7 Llama-3-70b 6 8 Llama-3-70b 9 Llama-3-70b 8 7 Llama-2-7b 8 7 Tie 10 Llama-3-70b 8 7 Llama-3-70b 8 9 Llama-3-70b 11 Llama-3-70b 8 7 Llama-2-7b 8 7 Tie 12 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 13 Llama-3-70b 8 6 Llama-2-7b 8 7 Tie 14 Llama-3-70b 8 7 Llama-2-7b 8 7 Tie 15 Llama-3-70b 9 7 Llama-2-7b 8 7 Tie 16 tie 8 8 Llama-3-70b 8 9 Llama-3-70b 17 Llama-2-7b 7 8 Llama-3-70b 7 9 Tie 18 Llama-3-70b 8 7 Llama-2-7b 8 7 Tie 19 Llama-3-70b 9 8 Llama-2-7b 9 7 Tie 20 Llama-2-7b 7 8 Llama-3-70b 6 8 Tie 21 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 22 Llama-2-7b 7 8 Llama-3-70b 6 8 Tie 23 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 24 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 25 Llama-3-70b 8 7 Llama-3-70b 8 9 Llama-3-70b C.5 Llama-2-7b vs. Vicuna-7b-v1.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-7b Vicuna-7b-v1.5 Winner Vicuna-7b-v1.5 Llama-2-7b Overall 1 Llama-2-7b 9 7 Vicuna-7b-v1.5 8 7 Tie 2 Vicuna-7b-v1.5 6 8 Llama-2-7b 7 8 Tie 3 Llama-2-7b 8 7 tie 8 8 Llama-2-7b 4 Llama-2-7b 8 7 Vicuna-7b-v1.5 8 7 Tie 5 Llama-2-7b 9 8 Llama-2-7b 8 9 Llama-2-7b 6 Llama-2-7b 8 7 Llama-2-7b 7 8 Llama-2-7b 7 Vicuna-7b-v1.5 7 8 Llama-2-7b 7 8 Tie 8 Vicuna-7b-v1.5 7 8 Llama-2-7b 7 8 Tie 9 Llama-2-7b 8 7 Vicuna-7b-v1.5 8 7 Tie 10 Llama-2-7b 8 7 Llama-2-7b 7 8 Llama-2-7b 11 Llama-2-7b 8 7 tie 8 8 Llama-2-7b 12 Vicuna-7b-v1.5 7 8 Llama-2-7b 6 8 Tie 13 Llama-2-7b 9 7 Llama-2-7b 7 8 Llama-2-7b 14 Llama-2-7b 8 7 Vicuna-7b-v1.5 8 7 Tie 15 Llama-2-7b 8 7 Vicuna-7b-v1.5 8 7 Tie 16 Llama-2-7b 7 6 Llama-2-7b 7 8 Llama-2-7b 17 Vicuna-7b-v1.5 6 8 Llama-2-7b 6 8 Tie 18 Llama-2-7b 9 7 Vicuna-7b-v1.5 8 7 Tie 19 Llama-2-7b 9 7 Llama-2-7b 8 9 Llama-2-7b 20 Llama-2-7b 8 7 Llama-2-7b 6 8 Llama-2-7b 21 Llama-2-7b 8 7 Llama-2-7b 8 9 Llama-2-7b 22 Vicuna-7b-v1.5 6 8 Llama-2-7b 5 7 Tie 23 Vicuna-7b-v1.5 6 8 Llama-2-7b 6 8 Tie 24 Llama-2-7b 8 7 Llama-2-7b 7 8 Llama-2-7b 25 Vicuna-7b-v1.5 7 8 Llama-2-7b 7 8 Tie 2051 C.6 Llama-2-7b vs. Vicuna-13b-v1.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-7b Vicuna-13b-v1.5 Winner Vicuna-13b-v1.5 Llama-2-7b Overall 1 Llama-2-7b 8 7 Vicuna-13b-v1.5 8 7 Tie 2 Vicuna-13b-v1.5 7 8 Llama-2-7b 7 8 Tie 3 Llama-2-7b 8 7 Vicuna-13b-v1.5 8 7 Tie 4 Llama-2-7b 8 7 Llama-2-7b 7 8 Llama-2-7b 5 Llama-2-7b 8 7 Vicuna-13b-v1.5 8 7 Tie 6 Llama-2-7b 8 7 Llama-2-7b 7 8 Llama-2-7b 7 Vicuna-13b-v1.5 8 9 Llama-2-7b 7 8 Tie 8 Vicuna-13b-v1.5 6 8 Llama-2-7b 7 8 Tie 9 Llama-2-7b 8 7 Vicuna-13b-v1.5 7 6 Tie 10 Llama-2-7b 8 7 Llama-2-7b 8 9 Llama-2-7b 11 Llama-2-7b 8 7 Vicuna-13b-v1.5 8 7 Tie 12 Vicuna-13b-v1.5 8 9 Llama-2-7b 6 9 Tie 13 Llama-2-7b 8 7 Llama-2-7b 7 8 Llama-2-7b 14 Llama-2-7b 8 7 Vicuna-13b-v1.5 8 7 Tie 15 Llama-2-7b 8 6 Llama-2-7b 7 8 Llama-2-7b 16 Vicuna-13b-v1.5 8 8 Llama-2-7b 7 8 Tie 17 Vicuna-13b-v1.5 7 8 Llama-2-7b 7 9 Tie 18 Llama-2-7b 8 7 Vicuna-13b-v1.5 8 7 Tie 19 Llama-2-7b 9 7 Llama-2-7b 8 9 Llama-2-7b 20 Vicuna-13b-v1.5 7 8 Llama-2-7b 6 8 Tie 21 Vicuna-13b-v1.5 8 9 Llama-2-7b 7 8 Tie 22 Vicuna-13b-v1.5 7 8 Llama-2-7b 7 8 Tie 23 Vicuna-13b-v1.5 7 8 Vicuna-13b-v1.5 8 7 Vicuna-13b-v1.5 24 Llama-2-7b 8 7 Llama-2-7b 6 8 Llama-2-7b 25 tie 8 8 Vicuna-13b-v1.5 8 7 Vicuna-13b-v1.5 C.7 Llama-2-7b vs. Mixtral-8x7B Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-7b Mixtral-8x7B Winner Mixtral-8x7B Llama-2-7b Overall 1 Llama-2-7b 8 7 Mixtral-8x7B 8 7 Tie 2 Mixtral-8x7B 7 8 Llama-2-7b 7 8 Tie 3 Llama-2-7b 8 7 Mixtral-8x7B 8 7 Tie 4 Llama-2-7b 8 7 Mixtral-8x7B 8 7 Tie 5 Llama-2-7b 8 7 Mixtral-8x7B 8 7 Tie 6 Llama-2-7b 8 7 Llama-2-7b 8 9 Llama-2-7b 7 Mixtral-8x7B 7 8 Llama-2-7b 7 8 Tie 8 Mixtral-8x7B 7 8 Llama-2-7b 7 8 Tie 9 Mixtral-8x7B 8 9 Draw 8 8 Mixtral-8x7B 10 Llama-2-7b 8 7 Mixtral-8x7B 8 7 Tie 11 Llama-2-7b 8 7 Mixtral-8x7B 8 7 Tie 12 Mixtral-8x7B 7 8 Llama-2-7b 8 9 Tie 13 Mixtral-8x7B 7 8 Mixtral-8x7B 8 7 Mixtral-8x7B 14 Mixtral-8x7B 7 8 Mixtral-8x7B 8 7 Mixtral-8x7B 15 Llama-2-7b 8 7 Mixtral-8x7B 8 7 Tie 16 Mixtral-8x7B 8 9 Mixtral-8x7B 8 7 Mixtral-8x7B 17 Mixtral-8x7B 7 8 Llama-2-7b 7 8 Tie 18 Llama-2-7b 8 7 Mixtral-8x7B 8 7 Tie 19 Llama-2-7b 8 7 Mixtral-8x7B 8 7 Tie 20 Mixtral-8x7B 7 8 Llama-2-7b 7 8 Tie 21 Mixtral-8x7B 8 9 Mixtral-8x7B 8 7 Mixtral-8x7B 22 Mixtral-8x7B 6 8 Llama-2-7b 7 8 Tie 23 Mixtral-8x7B 7 8 Mixtral-8x7B 8 7 Mixtral-8x7B 24 Mixtral-8x7B 7 8 Mixtral-8x7B 8 7 Mixtral-8x7B 25 Mixtral-8x7B 8 9 Mixtral-8x7B 8 7 Mixtral-8x7B 2052 C.8 Llama-2-7b vs GPT 3.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-7b GPT-3.5 Winner GPT-3.5 Llama-2-7b Overall 1 Llama-2-7b 8 7 GPT-3.5 9 7 Tie 2 GPT-3.5 7 8 Llama-2-7b 7 8 Tie 3 GPT-3.5 8 9 Llama-2-7b 8 9 Tie 4 Llama-2-7b 8 7 Llama-2-7b 7 8 Llama-2-7b 5 GPT-3.5 8 9 GPT-3.5 8 7 GPT-3.5 6 GPT-3.5 8 9 Llama-2-7b 7 8 Tie 7 GPT-3.5 7 8 Llama-2-7b 7 8 Tie 8 GPT-3.5 6 8 Llama-2-7b 8 7 Tie 9 Llama-2-7b 8 7 GPT-3.5 8 7 Tie 10 GPT-3.5 8 9 GPT-3.5 8 7 GPT-3.5 11 Llama-2-7b 8 7 GPT-3.5 8 7 Tie 12 GPT-3.5 7 9 Llama-2-7b 7 9 Tie 13 Llama-2-7b 8 7 GPT-3.5 8 7 Tie 14 GPT-3.5 8 9 GPT-3.5 8 7 GPT-3.5 15 Llama-2-7b 8 7 GPT-3.5 8 7 Tie 16 GPT-3.5 8 9 Llama-2-7b 8 9 Tie 17 GPT-3.5 7 9 Llama-2-7b 7 8 Tie 18 Llama-2-7b 8 7 GPT-3.5 8 7 Tie 19 Llama-2-7b 8 7 GPT-3.5 8 7 Tie 20 GPT-3.5 7 8 Llama-2-7b 7 8 Tie 21 GPT-3.5 8 9 GPT-3.5 8 7 GPT-3.5 22 GPT-3.5 6 8 Llama-2-7b 7 8 Tie 23 GPT-3.5 7 8 GPT-3.5 8 7 GPT-3.5 24 GPT-3.5 7 8 Llama-2-7b 7 8 Tie 25 GPT-3.5 8 9 Llama-2-7b 8 9 Tie C.9 Llama-2-7b vs GPT 4 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-7b GPT-4 Winner GPT-4 Llama-2-7b Overall 1 Llama-2-7b 9 8 GPT-4 9 7 Tie 2 GPT-4 7 9 GPT-4 8 7 GPT-4 3 GPT-4 8 9 GPT-4 8 7 GPT-4 4 GPT-4 8 9 GPT-4 8 7 GPT-4 5 GPT-4 8 9 GPT-4 9 8 GPT-4 6 GPT-4 7 9 GPT-4 8 7 GPT-4 7 GPT-4 7 9 GPT-4 8 7 GPT-4 8 GPT-4 6 8 GPT-4 8 7 GPT-4 9 GPT-4 7 8 GPT-4 9 7 GPT-4 10 GPT-4 7 8 GPT-4 8 7 GPT-4 11 GPT-4 7 8 GPT-4 8 7 GPT-4 12 GPT-4 6 9 GPT-4 8 7 GPT-4 13 GPT-4 7 8 GPT-4 9 7 GPT-4 14 GPT-4 7 9 GPT-4 9 7 GPT-4 15 GPT-4 7 8 GPT-4 9 7 GPT-4 16 GPT-4 7 9 GPT-4 8 7 GPT-4 17 GPT-4 7 9 GPT-4 8 7 GPT-4 18 GPT-4 8 9 GPT-4 9 7 GPT-4 19 GPT-4 8 9 GPT-4 8 7 GPT-4 20 GPT-4 6 8 GPT-4 8 7 GPT-4 21 GPT-4 7 9 GPT-4 8 7 GPT-4 22 GPT-4 7 9 GPT-4 8 7 GPT-4 23 GPT-4 7 8 GPT-4 8 7 GPT-4 24 GPT-4 7 9 GPT-4 8 7 GPT-4 25 GPT-4 8 9 GPT-4 8 7 GPT-4 2053 C.10 Llama-2-13b vs Llama-2-13b Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-13b Llama-2-13b Winner Llama-2-13b Llama-2-13b 1 Side 1 8 7 Side 1 8 7 2 Side 2 8 9 Side 2 7 8 3 Side 1 8 7 Side 1 8 7 4 Side 1 8 7 Side 1 8 7 5 Side 1 8 7 Side 1 8 7 6 Side 2 8 9 Side 1 8 7 7 Side 1 9 8 Side 1 8 7 8 Side 2 7 8 Side 2 7 8 9 Side 1 8 7 Side 1 8 7 10 Side 1 8 7 Side 1 8 7 11 Side 2 7 8 Side 1 8 7 12 Side 2 7 8 Side 2 7 8 13 Side 1 8 7 Side 1 8 7 14 Side 1 8 7 Side 1 8 7 15 Side 1 8 7 Side 1 8 7 16 Side 1 8 8 Side 1 8 8 17 Side 2 7 8 Side 2 7 8 18 Side 1 8 7 Side 1 8 7 19 Side 1 9 7 Side 1 9 7 20 Side 2 6 8 Side 2 7 8 21 Side 2 8 9 Side 2 8 9 22 Side 2 6 8 Side 2 6 8 23 Side 1 8 7 Side 1 8 7 24 Side 2 8 9 Side 2 8 9 25 Side 2 8 9 Side 2 8 9 C.11 Llama-2-13b vs Llama-2-70b Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-13b Llama-2-70b Winner Llama-2-70b Llama-2-13b Overall 1 Llama-2-13b 8 6 Llama-2-70b 8 7 Tie 2 Llama-2-70b 7 8 Llama-2-13b 8 9 Tie 3 Llama-2-70b 8 9 Llama-2-70b 8 7 Llama-2-70b 4 Llama-2-13b 8 7 Llama-2-70b 8 7 Tie 5 Llama-2-13b 8 7 Llama-2-70b 8 7 Tie 6 Llama-2-70b 8 9 Llama-2-70b 8 7 Llama-2-70b 7 Llama-2-70b 7 8 Llama-2-70b 8 7 Llama-2-70b 8 Llama-2-70b 7 8 Llama-2-70b 8 7 Llama-2-70b 9 Llama-2-70b 7 8 Llama-2-70b 8 7 Llama-2-70b 10 Llama-2-13b 8 7 Llama-2-70b 8 7 Tie 11 Llama-2-13b 8 7 Llama-2-70b 8 7 Tie 12 Llama-2-70b 7 8 Llama-2-13b 8 9 Tie 13 Llama-2-13b 8 7 Llama-2-70b 8 7 Tie 14 Llama-2-13b 8 7 Llama-2-70b 8 7 Tie 15 Llama-2-13b 8 6 Llama-2-70b 8 6 Tie 16 Llama-2-70b 8 9 Llama-2-70b 8 7 Llama-2-70b 17 Llama-2-70b 7 8 Llama-2-13b 7 8 Tie 18 Llama-2-13b 8 7 Llama-2-70b 8 7 Tie 19 Llama-2-13b 8 6 Llama-2-70b 8 6 Tie 20 Llama-2-70b 6 8 Llama-2-13b 7 8 Tie 21 Llama-2-70b 8 9 Llama-2-70b 8 7 Llama-2-70b 22 Llama-2-70b 6 8 Llama-2-13b 7 8 Tie 23 Llama-2-13b 8 7 Llama-2-70b 8 7 Tie 24 Llama-2-70b 7 8 Llama-2-13b 8 9 Tie 25 Llama-2-13b 8 7 Llama-2-70b 8 8 Tie 2054 C.12 Llama-2-13b vs. Llama-3-70b Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-3-70b Llama-2-13b Winner Llama-2-13b Llama-3-70b Overall 1 Llama-3-70b 9 7 Llama-2-13b 8 7 Tie 2 tie 8 8 Llama-3-70b 7 8 Llama-3-70b 3 Llama-3-70b 8 7 Llama-3-70b 8 9 Llama-3-70b 4 Llama-3-70b 8 7 Llama-2-13b 8 7 Tie 5 Llama-3-70b 8 7 Llama-2-13b 8 7 Tie 6 Llama-2-13b 8 9 Llama-2-13b 8 7 Llama-2-13b 7 Llama-3-70b 8 7 Llama-3-70b 8 9 Llama-3-70b 8 Llama-2-13b 7 8 Llama-3-70b 6 8 Tie 9 Llama-3-70b 8 7 Llama-3-70b 8 9 Llama-3-70b 10 Llama-3-70b 8 7 Llama-2-13b 8 7 Llama-2-13b 11 Llama-3-70b 8 6 Llama-3-70b 7 8 Llama-3-70b 12 Llama-3-70b 8 7 Llama-3-70b 7 9 Llama-3-70b 13 Llama-3-70b 8 6 Llama-3-70b 7 8 Llama-3-70b 14 Llama-3-70b 8 7 Llama-2-13b 8 7 Tie 15 Llama-3-70b 9 7 Llama-2-13b 8 7 Tie 16 Llama-3-70b 8 7 Llama-3-70b 8 9 Llama-3-70b 17 Llama-2-13b 7 8 Llama-3-70b 7 9 Tie 18 Llama-3-70b 9 8 Llama-3-70b 7 8 Llama-3-70b 19 Llama-3-70b 9 7 Llama-2-13b 8 7 Tie 20 Llama-2-13b 7 8 Llama-3-70b 6 8 Tie 21 Llama-3-70b 9 8 Llama-3-70b 7 8 Llama-3-70b 22 Llama-2-13b 6 8 Llama-3-70b 6 8 Ties 23 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 24 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 25 Llama-3-70b 8 7 Llama-3-70b 8 9 Llama-3-70b C.13 Llama-2-13b vs. Vicuna-7b-v1.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-13b Vicuna-7b-v1.5 Winner Vicuna-7b-v1.5 Llama-2-13b Overall 1 Llama-2-13b 8 7 Vicuna-7b-v1.5 8 7 Tie 2 Vicuna-7b-v1.5 7 8 Llama-2-13b 7 8 Tie 3 Llama-2-13b 9 8 Vicuna-7b-v1.5 8 8 Tie 4 Llama-2-13b 8 7 Vicuna-7b-v1.5 8 7 Tie 5 Llama-2-13b 8 7 Vicuna-7b-v1.5 8 7 Tie 6 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 7 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 8 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 9 Llama-2-13b 8 7 Vicuna-7b-v1.5 8 7 Tie 10 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 11 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 12 Llama-2-13b 8 7 Llama-2-13b 6 8 Llama-2-13b 13 Llama-2-13b 9 7 Llama-2-13b 7 8 Llama-2-13b 14 Llama-2-13b 8 7 Vicuna-7b-v1.5 8 7 Tie 15 Llama-2-13b 8 6 Vicuna-7b-v1.5 8 7 Tie 16 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 17 Llama-2-13b 8 7 Llama-2-13b 6 9 Llama-2-13b 18 Llama-2-13b 8 6 Vicuna-7b-v1.5 8 7 Tie 19 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 20 Vicuna-7b-v1.5 6 8 Llama-2-13b 6 8 Tie 21 Llama-2-13b 8 7 Llama-2-13b 6 8 Llama-2-13b 22 Vicuna-7b-v1.5 6 8 Llama-2-13b 6 8 Tie 23 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 24 Llama-2-13b 8 7 Llama-2-13b 8 9 Llama-2-13b 25 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 2055 C.14 Llama-2-13b vs. Vicuna-13b-v1.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-13b Vicuna-13b-v1.5 Winner Vicuna-13b-v1.5 Llama-2-13b Overall 1 Llama-2-13b 8 7 Vicuna-13b-v1.5 8 7 Tie 2 Vicuna-13b-v1.5 8 9 Llama-2-13b 7 8 Tie 3 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 4 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 5 Llama-2-13b 8 7 Vicuna-13b-v1.5 8 7 Tie 6 Vicuna-13b-v1.5 7 8 Llama-2-13b 7 8 Tie 7 Vicuna-13b-v1.5 8 9 Llama-2-13b 7 8 Tie 8 Vicuna-13b-v1.5 7 8 Llama-2-13b 6 8 Tie 9 Llama-2-13b 8 7 Vicuna-13b-v1.5 8 7 Tie 10 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 11 Llama-2-13b 8 7 Vicuna-13b-v1.5 8 7 Tie 12 Vicuna-13b-v1.5 8 9 Llama-2-13b 6 8 Tie 13 Llama-2-13b 8 6 Vicuna-13b-v1.5 8 7 Tie 14 Llama-2-13b 8 7 Vicuna-13b-v1.5 8 7 Tie 15 Llama-2-13b 8 6 Vicuna-13b-v1.5 8 7 Tie 16 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 17 Vicuna-13b-v1.5 7 8 Llama-2-13b 6 8 Tie 18 Llama-2-13b 8 7 Vicuna-13b-v1.5 8 7 Tie 19 Llama-2-13b 9 7 Llama-2-13b 7 8 Llama-2-13b 20 Vicuna-13b-v1.5 7 8 Llama-2-13b 6 8 Tie 21 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 22 Vicuna-13b-v1.5 6 8 Llama-2-13b 7 8 Tie 23 Llama-2-13b 8 7 Llama-2-13b 7 8 Llama-2-13b 24 Llama-2-13b 8 7 Vicuna-13b-v1.5 8 7 Tie 25 Llama-2-13b 8 7 Llama-2-13b 8 9 Llama-2-13b C.15 Llama-2-13b vs Mixtral-8x7B Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-13b Mixtral-8x7B Winner Mixtral-8x7B Llama-2-13b Overall 1 Llama-2-13b 8 7 Mixtral-8x7B 9 7 Tie 2 Mixtral-8x7B 7 8 Llama-2-13b 8 9 Tie 3 Llama-2-13b 8 7 Mixtral-8x7B 8 7 Tie 4 Mixtral-8x7B 8 9 Mixtral-8x7B 8 7 Mixtral-8x7B 5 Llama-2-13b 8 7 Mixtral-8x7B 8 7 Tie 6 Mixtral-8x7B 8 9 Llama-2-13b 7 8 Tie 7 Mixtral-8x7B 8 9 Llama-2-13b 7 8 Tie 8 Mixtral-8x7B 7 8 Llama-2-13b 6 8 Tie 9 Llama-2-13b 8 7 Mixtral-8x7B 8 7 Tie 10 Llama-2-13b 8 7 Mixtral-8x7B 8 7 Tie 11 Llama-2-13b 8 7 Mixtral-8x7B 8 7 Tie 12 Mixtral-8x7B 6 8 Llama-2-13b 8 9 Tie 13 Llama-2-13b 8 7 Mixtral-8x7B 8 7 Tie 14 Llama-2-13b 8 7 Mixtral-8x7B 8 7 Tie 15 Llama-2-13b 8 6 Mixtral-8x7B 8 7 Tie 16 Llama-2-13b 8 7 Mixtral-8x7B 8 7 Tie 17 Mixtral-8x7B 7 8 Llama-2-13b 7 8 Tie 18 Llama-2-13b 8 7 Mixtral-8x7B 8 7 Tie 19 Llama-2-13b 8 7 Mixtral-8x7B 8 7 Tie 20 Mixtral-8x7B 6 8 Llama-2-13b 6 9 Tie 21 Mixtral-8x7B 8 9 Mixtral-8x7B 8 7 Mixtral-8x7B 22 Mixtral-8x7B 6 8 Llama-2-13b 6 8 Tie 23 Mixtral-8x7B 7 8 Mixtral-8x7B 8 7 Mixtral-8x7B 24 Mixtral-8x7B 7 8 Mixtral-8x7B 8 7 Mixtral-8x7B 25 Mixtral-8x7B 7 8 Mixtral-8x7B 8 7 Mixtral-8x7B 2056 C.16 Llama-2-13b vs. GPT 3.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-13b GPT 3.5 Winner GPT 3.5 Llama-2-13b Overall 1 Llama-2-13b 8 7 GPT 3.5 9 7 Tie 2 GPT 3.5 7 9 Llama-2-13b 8 9 Tie 3 Llama-2-13b 8 7 GPT 3.5 8 7 Tie 4 Llama-2-13b 8 7 GPT 3.5 8 7 Tie 5 GPT 3.5 8 9 GPT 3.5 8 7 GPT 3.5 6 GPT 3.5 8 9 Llama-2-13b 7 8 Tie 7 GPT 3.5 8 9 Llama-2-13b 7 8 Tie 8 GPT 3.5 7 9 GPT 3.5 8 7 GPT 3.5 9 Llama-2-13b 8 7 GPT 3.5 8 7 Tie 10 Llama-2-13b 8 7 GPT 3.5 8 7 Tie 11 GPT 3.5 8 9 Llama-2-13b 7 8 Tie 12 GPT 3.5 8 9 Llama-2-13b 7 8 Tie 13 GPT 3.5 7 8 GPT 3.5 8 7 GPT 3.5 14 tie 8 8 GPT 3.5 8 7 GPT 3.5 15 Llama-2-13b 8 7 GPT 3.5 9 7 Tie 16 GPT 3.5 8 9 GPT 3.5 8 7 GPT 3.5 17 GPT 3.5 6 8 Llama-2-13b 8 9 Tie 18 Llama-2-13b 8 7 GPT 3.5 8 7 Tie 19 Llama-2-13b 8 7 GPT 3.5 8 7 Tie 20 GPT 3.5 6 8 Llama-2-13b 6 9 Tie 21 GPT 3.5 8 9 Llama-2-13b 7 8 Tie 22 GPT 3.5 6 8 Llama-2-13b 7 8 Tie 23 GPT 3.5 8 9 GPT 3.5 8 7 GPT 3.5 24 GPT 3.5 7 8 Llama-2-13b 8 9 Tie 25 GPT 3.5 7 8 GPT 3.5 8 7 GPT 3.5 C.17 Llama-2-13b vs. GPT 4 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-13b GPT-4 Winner GPT-4 Llama-2-13b Overall 1 GPT-4 8 9 GPT-4 9 6 GPT-4 2 GPT-4 6 9 Llama-2-13b 8 9 Tie 3 GPT-4 8 9 GPT-4 9 8 GPT-4 4 GPT-4 7 8 GPT-4 8 7 GPT-4 5 GPT-4 8 9 GPT-4 9 8 GPT-4 6 GPT-4 7 9 GPT-4 8 7 GPT-4 7 GPT-4 7 9 GPT-4 8 7 GPT-4 8 GPT-4 7 8 GPT-4 8 7 GPT-4 9 GPT-4 8 9 GPT-4 8 7 GPT-4 10 GPT-4 8 9 GPT-4 8 7 GPT-4 11 GPT-4 7 8 GPT-4 9 7 GPT-4 12 GPT-4 7 9 Llama-2-13b 8 9 Tie 13 GPT-4 7 8 GPT-4 8 7 GPT-4 14 GPT-4 7 9 GPT-4 9 8 GPT-4 15 GPT-4 8 9 GPT-4 9 7 GPT-4 16 GPT-4 8 9 GPT-4 8 7 GPT-4 17 GPT-4 6 9 GPT-4 8 7 GPT-4 18 GPT-4 8 9 GPT-4 9 7 GPT-4 19 GPT-4 8 9 GPT-4 8 7 GPT-4 20 GPT-4 6 9 Llama-2-13b 8 9 Tie 21 GPT-4 7 9 GPT-4 8 7 GPT-4 22 GPT-4 5 8 GPT-4 8 7 GPT-4 23 GPT-4 7 8 GPT-4 8 7 GPT-4 24 GPT-4 7 8 GPT-4 8 7 GPT-4 25 GPT-4 8 9 GPT-4 8 7 GPT-4 2057 C.18 Llama-2-70b vs. Llama-2-70b Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-70b Llama-2-70b Winner Llama-2-70b Llama-2-70b 1 1 9 7 1 8 7 2 2 7 8 2 7 8 3 2 8 9 2 8 9 4 1 8 7 1 8 7 5 1 8 7 1 8 7 6 2 8 9 2 8 9 7 2 8 9 2 8 9 8 2 8 9 1 8 7 9 1 9 7 1 8 7 10 2 8 9 2 8 9 11 1 8 7 1 8 7 12 2 8 9 2 8 9 13 1 8 7 1 9 7 14 1 8 7 1 8 7 15 1 8 7 1 9 7 16 1 8 7 1 9 8 17 2 7 8 2 8 9 18 1 8 7 1 8 7 19 1 8 7 1 8 7 20 2 6 8 2 7 8 21 1 8 7 2 8 9 22 2 7 8 2 7 8 23 1 8 7 1 8 7 24 1 8 7 1 8 7 25 2 8 9 2 8 9 C.19 Llama-2-70b vs. Llama-3-70b Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-3-70b Llama-2-70b Winner Llama-2-70b Llama-3-70b Overall 1 Llama-3-70b 9 7 Llama-2-70b 8 7 Tie 2 Llama-2-70b 8 9 Llama-3-70b 7 9 Tie 3 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 4 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 5 Llama-3-70b 8 7 Llama-3-70b 8 9 Llama-3-70b 6 Llama-2-70b 8 9 Llama-3-70b 7 8 Tie 7 Llama-2-70b 7 9 Llama-3-70b 7 8 Tie 8 Llama-2-70b 7 8 Llama-3-70b 7 8 Tie 9 Llama-3-70b 8 7 Llama-2-70b 8 7 Tie 10 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 11 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 12 Llama-2-70b 8 9 Llama-3-70b 7 8 Tie 13 Llama-3-70b 9 7 Llama-3-70b 8 9 Llama-3-70b 14 Llama-3-70b 8 7 Llama-2-70b 8 7 Tie 15 Llama-3-70b 9 6 Llama-2-70b 9 8 Tie 16 Llama-2-70b 8 9 Llama-2-70b 8 7 Tie 17 Llama-2-70b 7 8 Llama-3-70b 7 8 Tie 18 Llama-3-70b 8 7 Llama-2-70b 8 7 Tie 19 Llama-3-70b 8 7 Llama-2-70b 8 7 Tie 20 Llama-2-70b 8 9 Llama-3-70b 7 9 Tie 21 Tie 8 8 Llama-3-70b 8 9 Llama-3-70b 22 Llama-2-70b 7 8 Llama-3-70b 6 8 Tie 23 Llama-3-70b 8 7 Llama-2-70b 8 7 Tie 24 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 25 Llama-3-70b 8 7 Llama-3-70b 8 9 Llama-3-70b C.22 Llama-2-70b vs. Mixtral-8x7B Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-70b Mixtral-8x7B Winner Mixtral-8x7B Llama-2-70b Overall 1 Llama-2-70b 9 7 Mixtral-8x7B 8 7 Tie 2 Mixtral-8x7B 8 9 Llama-2-70b 7 8 Tie 3 Llama-2-70b 8 7 Mixtral-8x7B 8 7 Tie 4 Llama-2-70b 8 7 Mixtral-8x7B 8 7 Tie 5 Llama-2-70b 8 7 Mixtral-8x7B 8 7 Tie 6 Mixtral-8x7B 8 9 Llama-2-70b 8 9 Tie 7 Llama-2-70b 8 7 Llama-2-70b 8 9 Llama-2-70b 8 Mixtral-8x7B 8 9 Llama-2-70b 7 8 Tie 9 Llama-2-70b 8 7 Mixtral-8x7B 8 7 Tie 10 Llama-2-70b 8 7 Mixtral-8x7B 8 7 Tie 11 Llama-2-70b 8 7 Mixtral-8x7B 8 7 Tie 12 Mixtral-8x7B 8 9 Llama-2-70b 8 9 Tie 13 Mixtral-8x7B 7 8 Mixtral-8x7B 8 7 Mixtral-8x7B 14 tie 8 8 Mixtral-8x7B 8 7 Mixtral-8x7B 15 Llama-2-70b 8 7 Mixtral-8x7B 8 7 Mixtral-8x7B 16 Mixtral-8x7B 8 9 Llama-2-70b 8 9 Tie 17 Mixtral-8x7B 7 8 Llama-2-70b 7 8 Tie 18 Llama-2-70b 8 7 Mixtral-8x7B 8 7 Tie 19 Llama-2-70b 8 7 Mixtral-8x7B 8 7 Tie 20 Mixtral-8x7B 7 8 Llama-2-70b 7 9 Tie 21 tie 8 8 Llama-2-70b 8 9 Llama-2-70b 22 Mixtral-8x7B 6 8 Llama-2-70b 6 8 Tie 23 Llama-2-70b 8 7 Mixtral-8x7B 8 7 Tie 24 Llama-2-70b 8 7 Mixtral-8x7B 8 7 Tie 25 Mixtral-8x7B 8 9 Llama-2-70b 8 9 Tie C.23 Llama-2-70b vs. GPT 3.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-70b GPT 3.5 Winner GPT 3.5 Llama-2-70b Overall 1 GPT 3.5 7 9 Llama-2-70b 8 Tie 2 Llama-2-70b 8 7 GPT 3.5 8 Tie 3 Llama-2-70b 9 8 Llama-2-70b 7 Llama-2-70b 4 GPT 3.5 7 8 Llama-2-70b 7 Tie 5 GPT 3.5 8 9 GPT 3.5 9 GPT 3.5 6 Llama-2-70b 9 8 GPT 3.5 8 Tie 7 Llama-2-70b 9 8 GPT 3.5 8 Tie 8 Llama-2-70b 8 7 GPT 3.5 8 Tie 9 GPT 3.5 8 9 Llama-2-70b 7 Tie 10 GPT 3.5 7 8 GPT 3.5 8 GPT 3.5 11 GPT 3.5 7 8 Llama-2-70b 7 Tie 12 Llama-2-70b 8 7 GPT 3.5 8 Tie 13 GPT 3.5 7 8 Llama-2-70b 7 Tie 14 Llama-2-70b 9 8 GPT 3.5 9 Tie 15 GPT 3.5 7 8 Llama-2-70b 7 Tie 16 Llama-2-70b 9 8 GPT 3.5 8 Tie 17 Llama-2-70b 9 8 GPT 3.5 8 Tie 18 GPT 3.5 7 8 Llama-2-70b 7 Tie 19 GPT 3.5 7 8 Llama-2-70b 7 Tie 20 Llama-2-70b 8 6 GPT 3.5 8 Tie 21 Llama-2-70b 9 8 GPT 3.5 8 Tie 22 Llama-2-70b 8 7 GPT 3.5 8 Tie 23 Llama-2-70b 9 8 GPT 3.5 8 Tie 24 Llama-2-70b 9 8 GPT 3.5 8 Tie 25 Llama-2-70b 9 8 GPT 3.5 9 Tie 2060 C.24 Llama-2-70b vs. GPT 4 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-2-70b GPT 4 Winner GPT 4 Llama-2-70b Over 1 Llama-2-70b 8 7 GPT 4 9 7 Tie 2 GPT 4 8 9 Llama-2-70b 8 9 Tie 3 GPT 4 7 8 GPT 4 9 8 GPT 4 4 GPT 4 7 8 GPT 4 8 7 GPT 4 5 GPT 4 8 9 GPT 4 8 7 GPT 4 6 GPT 4 8 9 GPT 4 8 7 GPT 4 7 GPT 4 7 8 GPT 4 8 7 GPT 4 8 GPT 4 7 8 GPT 4 8 7 GPT 4 9 GPT 4 8 9 GPT 4 9 7 GPT 4 10 GPT 4 7 8 GPT 4 8 7 GPT 4 11 GPT 4 7 8 GPT 4 8 7 GPT 4 12 GPT 4 7 9 Llama-2-70b 7 9 Tie 13 GPT 4 7 8 GPT 4 9 7 GPT 4 14 GPT 4 8 9 GPT 4 9 7 GPT 4 15 GPT 4 8 9 GPT 4 8 7 GPT 4 16 GPT 4 8 9 Llama-2-70b 8 9 Tie 17 GPT 4 7 8 tie 8 8 GPT 4 18 GPT 4 8 9 GPT 4 8 7 GPT 4 19 GPT 4 8 9 GPT 4 8 7 GPT 4 20 GPT 4 7 9 GPT 4 8 7 GPT 4 21 GPT 4 7 9 GPT 4 8 7 GPT 4 22 GPT 4 6 9 GPT 4 8 7 GPT 4 23 GPT 4 8 9 GPT 4 8 7 GPT 4 24 GPT 4 7 9 GPT 4 8 7 GPT 4 25 GPT 4 8 9 GPT 4 9 8 GPT 4 C.25 Llama-3-70b vs. Vicuna-7b-v1.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-3-70b Vicuna-7b-v1.5 Winner Vicuna-7b-v1.5 Llama-3-70b Overall 1 Llama-3-70b 8 7 Llama-3-70b 4 7 Llama-3-70b 2 Llama-3-70b 8 7 Llama-3-70b 7 9 Llama-3-70b 3 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 4 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 5 Llama-3-70b 9 8 Llama-3-70b 8 9 Llama-3-70b 6 Llama-3-70b 8 7 Llama-3-70b 6 8 Llama-3-70b 7 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 8 Llama-3-70b 8 7 Llama-3-70b 6 8 Llama-3-70b 9 Llama-3-70b 9 7 Llama-3-70b 8 9 Llama-3-70b 10 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 11 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 12 Llama-3-70b 9 7 Llama-3-70b 6 9 Llama-3-70b 13 Llama-3-70b 9 7 Llama-3-70b 7 8 Llama-3-70b 14 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 15 Llama-3-70b 8 6 Llama-3-70b 7 8 Llama-3-70b 16 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 17 Vicuna-7b-v1.5 7 8 Llama-3-70b 6 9 Tie 18 Llama-3-70b 9 7 Llama-3-70b 7 8 Llama-3-70b 19 Llama-3-70b 8 7 Vicuna-7b-v1.5 8 7 Tie 20 Llama-3-70b 8 7 Llama-3-70b 6 9 Llama-3-70b 21 Llama-3-70b 8 7 Llama-3-70b 6 8 Llama-3-70b 22 Llama-3-70b 8 7 Llama-3-70b 6 9 Llama-3-70b 23 Llama-3-70b 8 6 Llama-3-70b 6 8 Llama-3-70b 24 Llama-3-70b 8 7 Llama-3-70b 7 9 Llama-3-70b 25 Llama-3-70b 8 7 Llama-3-70b 7 9 Llama-3-70b 1 Llama-3-70b 9 8 Mixtral-8x7B 8 7 Tie 2 Llama-3-70b 8 7 Llama-3-70b 8 9 Llama-3-70b 3 Llama-3-70b 8 7 Llama-3-70b 8 9 Llama-3-70b 4 Llama-3-70b 8 7 Mixtral-8x7B 8 7 Tie 5 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 6 Mixtral-8x7B 8 9 Llama-3-70b 8 9 Tie 7 Mixtral-8x7B 7 8 Llama-3-70b 7 8 Tie 8 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 9 Mixtral-8x7B 8 9 Llama-3-70b 8 9 Tie 10 Llama-3-70b 8 7 Llama-3-70b 8 9 Llama-3-70b 11 Llama-3-70b 8 7 Mixtral-8x7B 8 7 Tie 12 Mixtral-8x7B 7 9 Llama-3-70b 7 8 Tie 13 Llama-3-70b 8 7 Llama-3-70b 7 8 Llama-3-70b 14 Llama-3-70b 8 7 Mixtral-8x7B 8 7 Tie 15 Llama-3-70b 9 8 Mixtral-8x7B 8 7 Tie 16 Mixtral-8x7B 8 9 Llama-3-70b 8 9 Tie 17 Mixtral-8x7B 8 9 Llama-3-70b 7 8 Tie 18 Llama-3-70b 9 8 Mixtral-8x7B 8 7 Tie 19 Llama-3-70b 8 7 Mixtral-8x7B 8 7 Tie 20 Mixtral-8x7B 7 8 Llama-3-70b 6 8 Tie 21 Mixtral-8x7B 7 8 Llama-3-70b 8 9 Tie 22 Mixtral-8x7B 7 8 Llama-3-70b 6 9 Tie 23 Mixtral-8x7B 7 8 Llama-3-70b 8 9 Tie 24 Mixtral-8x7B 8 9 Llama-3-70b 7 8 Tie 25 Mixtral-8x7B 7 8 Llama-3-70b 7 8 Tie 2062 C.28 Llama-3-70b vs. GPT-3.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-3-70b GPT-3.5 Winner GPT-3.5 Llama-3-70b Overall 1 Llama-3-70b 8 7 GPT-3.5 8 Tie 2 GPT-3.5 8 9 Llama-3-70b 7 Tie 3 GPT-3.5 8 9 Llama-3-70b 8 Tie 4 Llama-3-70b 8 7 GPT-3.5 8 Tie 5 tie 8 8 Llama-3-70b 8 Llama-3-70b 6 GPT-3.5 7 8 Llama-3-70b 7 Tie 7 GPT-3.5 8 9 Llama-3-70b 7 Tie 8 GPT-3.5 7 8 Llama-3-70b 6 Tie 9 Llama-3-70b 8 7 GPT-3.5 8 Tie 10 GPT-3.5 8 9 Llama-3-70b 8 Tie 11 Llama-3-70b 8 7 GPT-3.5 8 Tie 12 GPT-3.5 8 9 Llama-3-70b 6 Tie 13 Llama-3-70b 8 7 GPT-3.5 8 Tie 14 Llama-3-70b 8 7 Llama-3-70b 7 Llama-3-70b 15 Llama-3-70b 9 8 GPT-3.5 8 Tie 16 Llama-3-70b 9 8 Llama-3-70b 7 Llama-3-70b 17 GPT-3.5 8 9 Llama-3-70b 7 Tie 18 Llama-3-70b 8 7 Llama-3-70b 8 Llama-3-70b 19 Llama-3-70b 8 7 GPT-3.5 8 Tie 20 GPT-3.5 7 8 Llama-3-70b 7 Tie 21 GPT-3.5 8 9 Llama-3-70b 7 Tie 22 GPT-3.5 7 9 Llama-3-70b 6 Tie 23 Llama-3-70b 8 7 Llama-3-70b 7 Llama-3-70b 24 GPT-3.5 8 9 Llama-3-70b 7 Tie 25 GPT-3.5 8 9 Llama-3-70b 8 Tie C.29 Llama-3-70b vs. GPT-4 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama-3-70b GPT-4 Winner GPT-4 Llama-3-70b Overall 1 Llama-3-70b 8 7 GPT-4 9 Tie 2 GPT-4 7 9 Llama-3-70b 7 Tie 3 GPT-4 8 9 GPT-4 8 GPT-4 4 GPT-4 7 8 GPT-4 8 GPT-4 5 GPT-4 8 9 GPT-4 8 GPT-4 6 GPT-4 7 8 Llama-3-70b 8 Tie 7 GPT-4 8 9 GPT-4 8 GPT-4 8 GPT-4 7 8 GPT-4 8 GPT-4 9 GPT-4 8 9 GPT-4 8 GPT-4 10 GPT-4 8 9 Llama-3-70b 8 Tie 11 Llama-3-70b 8 7 GPT-4 8 Tie 12 GPT-4 7 8 Llama-3-70b 8 Tie 13 Llama-3-70b 8 7 GPT-4 8 Tie 14 GPT-4 8 9 GPT-4 9 GPT-4 15 GPT-4 8 9 GPT-4 9 GPT-4 16 GPT-4 8 9 Llama-3-70b 8 Tie 17 GPT-4 8 9 GPT-4 8 GPT-4 18 GPT-4 8 9 GPT-4 8 GPT-4 19 Llama-3-70b 8 7 GPT-4 9 Tie 20 GPT-4 7 9 Llama-3-70b 7 Tie 21 GPT-4 7 8 Llama-3-70b 8 Tie 22 GPT-4 7 9 GPT-4 8 GPT-4 23 GPT-4 7 9 GPT-4 8 GPT-4 24 GPT-4 8 9 Llama-3-70b 8 Tie 25 GPT-4 8 9 GPT-4 8 GPT-4 2063 C.30 Vicuna-7b-v1.5 vs. Vicuna-7b-v1.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Vicuna-7b-v1.5 Vicuna-7b-v1.5 Winner Vicuna-7b-v1.5 Vicuna-7b-v1.5 1 1 8 7 1 8 2 2 7 8 2 7 3 1 8 7 1 8 4 1 8 7 1 8 5 1 8 7 1 8 6 1 8 7 2 7 7 1 8 7 2 8 8 2 7 8 2 7 9 1 8 7 1 8 10 1 8 7 1 8 11 1 8 7 1 8 12 2 6 8 2 7 13 2 7 8 2 7 14 2 7 8 1 8 15 1 8 7 1 8 16 1 8 7 1 8 17 2 8 9 2 7 18 1 8 7 1 8 19 1 8 7 1 8 20 2 7 8 2 6 21 1 8 7 1 8 22 2 7 8 2 6 23 1 8 7 1 8 24 1 8 7 1 8 25 2 8 9 2 8 C.31 Vicuna-7b-v1.5 vs. Vicuna-13b-v1.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Vicuna-7b-v1.5 GPT 4 Winner GPT 4 Vicuna-7b-v1.5 Overall 1 GPT 4 8 9 GPT 4 9 7 GPT 4 2 GPT 4 7 9 GPT 4 9 7 GPT 4 3 GPT 4 8 9 GPT 4 9 7 GPT 4 4 GPT 4 7 9 GPT 4 9 8 GPT 4 5 GPT 4 7 9 GPT 4 9 8 GPT 4 6 GPT 4 7 9 GPT 4 9 7 GPT 4 7 GPT 4 7 9 Vicuna-7b-v1.5 8 9 Tie 8 GPT 4 6 8 GPT 4 8 7 GPT 4 9 GPT 4 7 8 GPT 4 9 7 GPT 4 10 GPT 4 7 8 GPT 4 8 7 GPT 4 11 GPT 4 7 9 GPT 4 8 7 GPT 4 12 GPT 4 6 9 GPT 4 8 7 GPT 4 13 GPT 4 7 8 GPT 4 9 7 GPT 4 14 GPT 4 7 8 GPT 4 9 8 GPT 4 15 GPT 4 7 9 GPT 4 9 6 GPT 4 16 GPT 4 7 9 GPT 4 9 8 GPT 4 17 GPT 4 7 9 GPT 4 8 7 GPT 4 18 GPT 4 8 9 GPT 4 9 7 GPT 4 19 GPT 4 8 9 GPT 4 9 8 GPT 4 20 GPT 4 6 9 GPT 4 8 7 GPT 4 21 GPT 4 7 8 GPT 4 8 7 GPT 4 22 GPT 4 6 9 GPT 4 8 7 GPT 4 23 GPT 4 7 9 GPT 4 9 8 GPT 4 24 GPT 4 7 9 GPT 4 9 7 GPT 4 25 GPT 4 7 9 GPT 4 8 7 GPT 4 C.35 Vicuna-13b-v1.5 vs. Vicuna-13b-v1.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Vicuna-13b-v1.5 Vicuna-13b-v1.5 Winner Vicuna-13b-v1.5 Vicuna-13b-v1.5 1 1 9 7 1 9 7 2 2 7 8 2 7 8 3 1 8 7 1 8 7 4 1 8 7 1 8 7 5 1 8 7 1 8 7 6 2 8 9 2 7 8 7 2 7 8 2 7 8 8 2 7 8 2 7 8 9 1 8 7 1 8 7 10 1 8 7 1 8 7 11 1 8 7 1 8 7 12 2 7 9 2 6 8 13 1 8 7 1 8 7 14 1 8 7 1 8 7 15 1 8 7 1 8 7 16 1 8 7 1 8 7 17 2 7 8 2 7 8 18 1 8 7 1 8 7 19 1 8 7 1 8 6 20 2 6 8 2 6 8 21 1 8 7 1 8 7 22 2 6 7 2 7 8 23 1 8 7 1 8 7 24 2 7 8 2 7 8 25 2 8 9 1 8 8 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Vicuna-13b-v1.5 GPT 4 Winner GPT 4 Vicuna-13b-v1.5 Overall 1 GPT 4 8 9 GPT 4 9 7 GPT 4 2 GPT 4 6 9 Vicuna-13b-v1.5 8 9 Tie 3 GPT 4 7 8 GPT 4 8 7 GPT 4 4 GPT 4 7 8 GPT 4 8 7 GPT 4 5 GPT 4 7 8 GPT 4 9 8 GPT 4 6 GPT 4 7 9 tie 8 8 GPT 4 7 GPT 4 7 9 Vicuna-13b-v1.5 8 9 Tie 8 GPT 4 6 9 GPT 4 8 7 GPT 4 9 GPT 4 8 9 GPT 4 9 7 GPT 4 10 GPT 4 7 9 GPT 4 8 7 GPT 4 11 GPT 4 7 8 GPT 4 9 7 GPT 4 12 GPT 4 6 9 GPT 4 9 8 GPT 4 13 GPT 4 7 8 GPT 4 9 7 GPT 4 14 GPT 4 7 9 GPT 4 9 7 GPT 4 15 GPT 4 7 8 GPT 4 9 7 GPT 4 16 GPT 4 7 9 GPT 4 9 8 GPT 4 17 GPT 4 6 8 GPT 4 8 7 GPT 4 18 GPT 4 8 9 GPT 4 9 8 GPT 4 19 GPT 4 7 8 GPT 4 9 7 GPT 4 20 GPT 4 6 9 GPT 4 8 7 GPT 4 21 GPT 4 7 9 GPT 4 8 7 GPT 4 22 GPT 4 6 8 GPT 4 8 7 GPT 4 23 GPT 4 7 9 GPT 4 8 7 GPT 4 24 GPT 4 7 9 GPT 4 8 7 GPT 4 25 GPT 4 7 9 tie 8 8 GPT 4 C.39 Mixtral-8x7B vs. Mixtral-8x7B Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Mixtral-8x7B Mixtral-8x7B Winner Mixtral-8x7B Mixtral-8x7B C.40 Mixtral-8x7B vs. GPT 3.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Mixtral-8x7B GPT 3.5 Winner GPT 3.5 Mixtral-8x7B Overall 1 Mixtral-8x7B 8 7 GPT 3.5 8 7 Tie 2 GPT 3.5 7 9 Mixtral-8x7B 7 8 Tie 3 Mixtral-8x7B 8 7 Mixtral-8x7B 8 9 Mixtral-8x7B 4 Mixtral-8x7B 8 7 GPT 3.5 8 7 Tie 5 GPT 3.5 8 9 Mixtral-8x7B 8 9 Tie 6 GPT 3.5 8 9 Mixtral-8x7B 8 9 Tie 7 GPT 3.5 8 9 Mixtral-8x7B 8 9 Tie 8 GPT 3.5 7 8 Mixtral-8x7B 7 8 Tie 9 Mixtral-8x7B 8 7 GPT 3.5 8 7 Tie 10 GPT 3.5 8 9 Mixtral-8x7B 8 9 Tie 11 Mixtral-8x7B 8 7 GPT 3.5 8 7 Tie 12 GPT 3.5 7 9 Mixtral-8x7B 7 8 Tie 13 GPT 3.5 7 8 GPT 3.5 8 7 Mixtral-8x7B 14 GPT 3.5 8 9 GPT 3.5 8 8 Mixtral-8x7B 15 Mixtral-8x7B 8 7 GPT 3.5 8 7 Tie 16 GPT 3.5 8 9 Mixtral-8x7B 8 9 Tie 17 GPT 3.5 7 8 Mixtral-8x7B 7 9 Tie 18 Mixtral-8x7B 8 7 GPT 3.5 8 7 Tie 19 Mixtral-8x7B 8 7 GPT 3.5 8 7 Tie 20 GPT 3.5 7 9 Mixtral-8x7B 7 8 Tie 21 GPT 3.5 7 8 Mixtral-8x7B 8 9 Tie 22 GPT 3.5 7 8 Mixtral-8x7B 7 8 Tie 23 tie 8 8 GPT 3.5 8 7 Mixtral-8x7B 24 GPT 3.5 8 9 Mixtral-8x7B 7 8 Tie 25 GPT 3.5 8 9 Mixtral-8x7B 8 9 Tie C.41 Mixtral-8x7B vs. GPT 4 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Mixtral-8x7B GPT 4 Winner GPT 4 Mixtral-8x7B Overall 1 GPT 4 8 9 GPT 4 9 7 GPT 4 2 GPT 4 6 9 Mixtral-8x7B 8 9 Tie 3 GPT 4 7 8 GPT 4 8 7 GPT 4 4 GPT 4 7 8 GPT 4 8 7 GPT 4 5 GPT 4 7 8 GPT 4 9 8 GPT 4 6 GPT 4 7 9 tie 8 8 GPT 4 7 GPT 4 7 9 Mixtral-8x7B 8 9 Tie 8 GPT 4 6 9 GPT 4 8 7 GPT 4 9 GPT 4 8 9 GPT 4 9 7 GPT 4 10 GPT 4 7 9 GPT 4 8 7 GPT 4 11 GPT 4 7 8 GPT 4 9 7 GPT 4 12 GPT 4 6 9 GPT 4 9 8 GPT 4 13 GPT 4 7 8 GPT 4 9 7 GPT 4 14 GPT 4 7 9 GPT 4 9 7 GPT 4 15 GPT 4 7 8 GPT 4 9 7 GPT 4 16 GPT 4 7 9 GPT 4 9 8 GPT 4 17 GPT 4 6 8 GPT 4 8 7 GPT 4 18 GPT 4 8 9 GPT 4 9 8 GPT 4 19 GPT 4 7 8 GPT 4 9 7 GPT 4 20 GPT 4 6 9 GPT 4 8 7 GPT 4 21 GPT 4 7 9 GPT 4 8 7 GPT 4 22 GPT 4 6 8 GPT 4 8 7 GPT 4 23 GPT 4 7 9 GPT 4 8 7 GPT 4 24 GPT 4 7 9 GPT 4 8 7 GPT 4 25 GPT 4 7 9 tie 8 8 GPT 4 2069 C.42 GPT 3.5 vs. GPT 3.5 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner GPT 3.5 GPT 3.5 Winner GPT 3.5 GPT 3.5 1 1 9 8 1 8 7 2 2 7 9 2 8 9 3 2 8 9 2 8 9 4 2 7 8 1 8 7 5 2 8 9 2 8 9 6 2 8 9 2 7 8 7 2 7 8 2 7 8 8 2 7 8 2 7 8 9 1 8 7 2 8 9 10 1 8 7 2 8 9 11 1 8 7 2 8 9 12 2 7 8 2 7 8 13 1 8 7 1 8 7 14 2 8 9 2 8 9 15 1 8 7 1 8 7 16 2 8 9 1 8 7 17 2 7 8 2 7 8 18 1 8 7 1 8 8 19 1 8 7 1 8 7 20 2 7 9 2 7 9 21 2 7 8 2 8 9 22 2 7 8 2 6 8 23 1 8 7 1 8 7 24 2 8 9 2 8 9 25 2 8 9 2 7 8 C.43 GPT 3.5 vs. GPT 4 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner GPT 3.5 GPT 4 Winner GPT 4 GPT 3.5 Overall 1 GPT 4 8 9 GPT 4 9 GPT 4 2 GPT 4 7 9 GPT 3.5 8 Tie 3 GPT 4 8 9 GPT 4 9 GPT 4 4 GPT 4 7 8 GPT 4 8 GPT 4 5 GPT 4 8 9 GPT 4 8 GPT 4 6 GPT 4 7 9 GPT 3.5 8 Tie 7 GPT 4 7 9 GPT 3.5 8 Tie 8 GPT 4 7 8 GPT 3.5 7 Tie 9 GPT 4 8 9 GPT 4 9 GPT 4 10 GPT 4 8 9 GPT 4 9 GPT 4 11 GPT 4 8 9 GPT 4 9 GPT 4 12 GPT 4 6 9 GPT 3.5 8 Tie 13 GPT 4 7 8 GPT 4 8 GPT 4 14 GPT 4 8 9 GPT 4 9 GPT 4 15 GPT 4 8 9 GPT 4 9 GPT 4 16 GPT 4 7 9 GPT 4 9 GPT 4 17 GPT 4 7 9 GPT 3.5 8 Tie 18 GPT 4 8 9 GPT 4 9 GPT 4 19 GPT 3.5 8 7 GPT 3.5 9 GPT 3.5 20 GPT 4 7 9 GPT 3.5 7 Tie 21 GPT 4 7 8 GPT 3.5 8 Tie 22 GPT 4 6 8 GPT 3.5 8 Tie 23 GPT 4 8 9 GPT 4 8 GPT 4 24 GPT 4 7 8 GPT 3.5 8 Tie 25 GPT 4 7 8 GPT 4 8 GPT 4 2070 C.44 GPT 4 vs. GPT 4 Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner GPT 4 GPT 4 Winner GPT 4 GPT 4 1 1 8 8 1 8 7 2 2 7 8 2 8 9 3 2 8 9 2 8 9 4 1 8 7 1 8 7 5 2 8 9 1 9 8 6 2 7 8 2 8 9 7 2 8 9 2 7 8 8 2 7 8 2 7 8 9 2 8 9 1 8 7 10 2 8 9 2 8 9 11 1 8 7 1 8 7 12 2 8 9 2 8 9 13 1 8 7 1 8 7 14 1 8 7 2 8 9 15 1 8 7 1 9 8 16 2 8 9 2 8 9 17 2 8 9 2 7 8 18 1 9 8 1 8 7 19 1 8 7 1 9 8 20 2 8 9 2 8 9 21 2 8 9 2 8 9 22 2 8 9 2 8 9 23 1 9 8 2 8 9 24 2 8 9 2 8 9 25 2 8 9 2 8 9
In this section, we report the results of experiment from Section 4.2 with Llama-3-70b as judge.
D.1 Llama2-13b vs. Llama2-7b
Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama2-7b Llama2-13b Winner Llama2-13b Llama2-7b Overall 1 Llama2-7b 8 7 Llama2-13b 8 7 Tie 2 Llama2-13b 8 9 Llama2-7b 8 9 Tie 3 Llama2-7b 8 7 Llama2-13b 8 7 Tie 4 Llama2-7b 8 6 Llama2-13b 8 6 Tie 5 Llama2-7b 8 7 Llama2-13b 8 7 Tie 6 Llama2-7b 8 7 Llama2-13b 8 7 Tie 7 Llama2-13b 8 9 Llama2-7b 8 9 Tie 8 Llama2-13b 8 9 Llama2-7b 8 9 Tie 9 Llama2-7b 8 7 Llama2-13b 8 7 Tie 10 Llama2-13b 8 9 Llama2-13b 8 7 Llama2-13b 11 Llama2-7b 8 7 Llama2-13b 8 6 Tie 12 Llama2-13b 8 9 Llama2-13b 8 7 Llama2-13b 13 Llama2-7b 8 7 Llama2-13b 8 6 Tie 14 Llama2-7b 8 7 Llama2-13b 8 7 Tie 15 Llama2-7b 8 6 Llama2-13b 8 6 Tie 16 Llama2-7b 8 7 Llama2-13b 8 7 Tie 17 Llama2-13b 8 9 Llama2-13b 8 7 Llama2-13b 18 Llama2-7b 8 7 Llama2-13b 8 7 Tie 19 Llama2-7b 8 7 Llama2-13b 8 6 Tie 20 Llama2-13b 7 8 Llama2-7b 8 9 Tie 21 Llama2-13b 8 9 Llama2-13b 8 7 Llama2-13b 22 Llama2-13b 7 8 Llama2-7b 6 8 Llama2-13b 23 Llama2-13b 8 9 Llama2-13b 8 7 Llama2-13b 24 Llama2-13b 8 9 Llama2-13b 8 7 Llama2-13b 25 Llama2-7b 8 7 Llama2-13b 8 7 Tie 2071 D.2 Llama2-13b vs. Llama2-70b Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama2-13b Llama2-70b Winner Llama2-70b Llama2-13b Overall 1 Llama2-13b 8 7 Llama2-70b 8 Tie 2 Llama2-70b 7 8 Llama2-70b 8 Llama2-70b 3 Llama2-70b 8 9 Llama2-13b 8 Tie 4 Llama2-13b 8 7 Llama2-70b 8 Tie 5 Llama2-13b 8 7 Llama2-70b 8 Tie 6 Llama2-70b 8 9 Llama2-70b 8 Llama2-70b 7 Llama2-13b 8 7 Llama2-70b 8 Tie 8 Llama2-13b 8 7 Llama2-70b 8 Tie 9 Llama2-13b 8 7 Llama2-70b 8 Tie 10 Llama2-70b 8 9 Llama2-13b 8 Tie 11 Llama2-13b 8 7 Llama2-70b 8 Tie 12 Llama2-70b 8 9 Llama2-13b 8 Tie 13 Llama2-13b 8 7 Llama2-70b 8 Tie 14 Llama2-13b 8 7 Llama2-70b 8 Tie 15 Llama2-13b 8 6 Llama2-70b 8 Tie 16 Llama2-70b 7 8 Llama2-70b 8 Llama2-70b 17 Llama2-13b 8 6 Llama2-70b 8 Tie 18 Llama2-13b 8 7 Llama2-70b 8 Tie 19 Llama2-13b 8 6 Llama2-70b 8 Tie 20 Llama2-70b 7 8 Llama2-13b 8 Tie 21 Llama2-13b 8 7 Llama2-70b 8 Tie 22 Llama2-70b 6 8 Llama2-13b 7 Tie 23 Llama2-13b 8 7 Llama2-70b 8 Tie 24 Llama2-70b 8 9 Llama2-13b 8 Tie 25 Llama2-70b 8 9 Llama2-70b 8 Llama2-70b D.3 Llama2-13b vs. Llama3-70b Home Away Side 1 Side 2 Side 1 Side 2 Topic Winner Llama3-70b Llama2-13b Winner Llama2-13b Llama3-70b Overall 1 Llama3-70b 8 7 Llama2-13b 8 Tie 2 Llama2-13b 8 9 Llama3-70b 8 Tie 3 Llama3-70b 8 7 Llama2-13b 8 Tie 4 Llama3-70b 8 6 Llama2-13b 8 Tie 5 Llama3-70b 8 7 Llama2-13b 8 Tie 6 Llama2-13b 8 9 Llama2-13b 8 Llama2-13b 7 Llama2-13b 8 9 Llama3-70b 8 Tie 8 Llama2-13b 8 9 Llama3-70b 8 Tie 9 Llama3-70b 8 7 Llama2-13b 8 Tie 10 Llama2-13b 8 9 Llama3-70b 7 Tie 11 Llama3-70b 8 7 Llama2-13b 8 Tie 12 Llama2-13b 8 9 Llama3-70b 8 Tie 13 Llama3-70b 8 7 Llama2-13b 8 Tie 14 Llama3-70b 8 7 Llama3-70b 8 Llama3-70b 15 Llama3-70b 8 7 Llama2-13b 8 Tie 16 Llama2-13b 8 9 Llama2-13b 8 Llama2-13b 17 Llama2-13b 8 9 Llama3-70b 7 Tie 18 Llama3-70b 8 7 Llama2-13b 8 Tie 19 Llama3-70b 8 7 Llama2-13b 8 Tie 20 Llama3-70b 8 7 Llama3-70b 8 Llama3-70b 21 Llama3-70b 8 7 Llama3-70b 8 Llama3-70b 22 Llama2-13b 8 9 Llama3-70b 7 Tie 23 Llama3-70b 8 7 Llama3-70b 8 Llama3-70b 24 Llama3-70b 8 7 Llama2-13b 8 Tie 25 Llama2-13b 8 9 Llama3-70b 8 Tie