In this work, we employ advanced prompt engineering techniques with pre-trained Large Language Models (LLMs) to enhance knowledge extraction and structuring for Knowledge Base Construction (KBC) tasks. At the core of our methodology is the strategic fusion of diferent prompting strategies tailored to the unique relations between subject and object entities. We choose diferent prompting strategies to overcome the challenges of null values, numeric data, and one-to-many relationships inherent in traditional KBC tasks. By integrating role-play and context-aware prompting, we enrich the interaction with the LLM, guiding it to produce more accurate and contextually relevant outputs. Our method achieved a macro-averaged F1-score of 0.653 across the properties, with the scores varying from 0.890 to 0.399. Our results show a marked improvement in precision and recall of the extracted data, highlighting the eficacy of our approach in transforming raw LLM outputs into structured, queryable knowledge bases. Our code base is publicly available for research and development purposes, accessible at: https://github.com/nobretincheva/challenge24.git
The emergence of advanced contextual Large Language Models (LLMs), exemplified by the
encoder-only models (e.g. BERT), decoder-only models (e.g. GPT, Mistral, Llama series), and
encoder-decoder (e.g. Flan-T5) models, has revealed a remarkable capacity for encapsulating
vast amounts of factual world knowledge within their parametric structures based on their
training methods and datasets. This internal representation of knowledge has been likened to
the schema-based relational knowledge base (KB) [
The ISWC KBC-LLM challenge seeks to advance the field of KBC through the exploration of LLM’s internal knowledge representations. Unlike previous work that focuses primarily on knowledge probing, this challenge involves the completion of factual, disambiguated knowledge bases for a given set of relations. Participants complete knowledge graphs by extracting entities and relations from LLMs, handling complexities such as null values, one-to-many relations, and cardinality constraints.
The ISWC’24 KBC-LLM challenge introduces constraints on model size (< = 10b parameters) and relation sets (=5 only) to foster innovation and fair competition. This challenge extends beyond previous knowledge extraction eforts by requiring the construction of comprehensive, disambiguated knowledge bases from LLMs. The task set by this challenge is to predict the object entities (null, zero, single, multiple) given the subject entity and the relation that is sourced from Wikidata. Given a subject entity and a specified relation, the task involves predicting a set of potential object entities using an LLM. Subsequently, these predictions undergo disambiguation and mapping to external knowledge bases to yield precise entity identifiers. This process necessitates handling diverse relational complexities, including null values, one-to-many relationships, and the extraction of multiple objects per query.
The intricate relational structures inherent within the challenge dataset render traditional knowledge probing methodologies insuficient. This work introduces a novel framework for fusion prompt strategy, incorporating dual, direct, and loop prompting methods tailored to the specific relational characteristics of the target knowledge, as well as context-aware prompting and role-play prompting. Due to the constraints imposed upon model size in this challenge we choose to use the Llama-3-8b-Instruct model. Using our fusion prompt strategy, we are able to achieve a macro-average F1 score of 0.653 on the test set, with F1-scores ranging from 0.399 in the awardWonBy relation to 0.890 for countryLandBordersCountry relation.
This section focuses on two key areas relevant to our tasks related to the ISWC Challenge: Knowledge Probing in LLMs and KB completion/construction. We highlight the state-of-the-art approaches and identify existing gaps in these domains.
Knowledge Probing in LLMs: The LAMA (LAnguage Model Analysis) probe [
Knowledge Base Completion with LLMs Traditional methods for knowledge base
completion primarily relied on three core approaches [
The use of LLMs as an alternative for KBC tasks, also known as text-based KBC methods, is
a relatively new area of research [
Our work’s main contribution is the fusion of prompt engineering techniques to address taskspecific challenges and efectively direct an LLM’s attention to the most relevant information within its training data. Employing a pre-trained LLM to predict the object entity given a subject entity and a relation entails evaluating the extent of knowledge captured by pre-trained LLMs during their training. To achieve that, we propose a fusion of various prompt engineering techniques. An overview of our methodology is illustrated in figure 1 and figure 2.
Input: SubjectEntity, SubjectEntityID, Relation
SubjectEntity,
Relation
Context Retrieval Role-play Persona Generation
CSuoRbnojteelecx-ttpEulanalty++iItnyP,feoRrrsmeolaanttaiioonn Prompt Generation
Prompt
Response Re-prompting Response DisamEbnitgituyation Llama-3-8B
Output: ObjectEntityID
The task is limited to five relations: countryLandBordersCountry, personHasCityOfDeath, seriesHasNumberOfEpisodes, awardWonBy, and companyTradesAtStockExchange. Some of these relations present a challenge in that sometimes there are no object entities i.e. null values may be present, such as in the following cases: • countryLandBordersCountry: Null values are possible (e.g., an island like Iceland). • personHasCityOfDeath: Null values are possible (e.g., the person is still alive). • awardWonBy: There may be instances where the award was not given in certain years. • companyTradesAtStockExchange: Null values are possible (e.g., private companies or have delisted from the stock exchange).
To address the challenge of null values, we developed a dual prompting approach known as
two-stage prompting. Dual prompting forms the cornerstone of our methodology, employing a
strategic two-tiered question-based prompting system that significantly enhances the LLM’s
ability to extract accurate knowledge [
The dual prompting method involves two stages: an initial prompt and a follow-up prompt. The initial prompt serves as a cognitive anchor, setting the context for the query and focusing the model’s attention on the presence or relevance of specific information. This is particularly useful in clarifying the existence or state of an entity, a crucial step in ensuring that subsequent inquiries are grounded in the correct context. The follow-up prompt then capitalizes on the clarified context to extract specific details or additional layers of information. Examples of these prompts are shown in table 6, and the prompt template can be found in table 7 and table 8. Our prompt templates are enhanced with instructions for the LLM on approaching and answering the question, specifying whether the answer should be yes/no, numeric, or multiple responses.
For the seriesHasNumberOfEpisodes relation, we use a single prompting approach, where the model is asked the question directly without any verification steps or follow-up prompts, as shown in figure 7. A null check is not necessary in this case because every series inherently has a defined number of episodes, eliminating the possibility of null values.
The execution strategy for prompting in our system employs tailored approaches, categorized as either looping or direct, depending on the nature of the relation involved. The strategies are applied as follows:
The looping strategy involves repeatedly applying the dual prompting method to retrieve relevant information for each instance and then aggregating the results to form the final answer. This strategy specifically addresses the challenge of the awardWonBy relation, where many objects are associated with a single subject that has multiple instances (e.g., 224 Physics Nobel Prize winners over the course of multiple years). This approach simplifies the LLM’s work and ensures accuracy by breaking the task into smaller, manageable parts. For the awardWonBy relation, the loop starts from the first year the award was conferred - to initialise this we prompt the LLM for said year. Upon entering the loop, for each subsequent year up to the present (2024), the LLM is prompted to list the award recipients. The final output is a combined list of all recipients across the years, ensuring a comprehensive and precise result. The strategy is illustrated in figure 3.
This strategy involves prompting the LLM directly to generate the object entity. We directly apply dual prompting to generate object entities for the relations countryLandBordersCountry, personHasCityOfDeath, and companyTradesAtStockExchange. We also use the direct strategy for the seriesHasNumberOfEpisodes relation; we directly prompt the total number of episodes without any dual prompting since we do not need to verify the series’ existence. For example, we might ask, "Final answer should consist of a number. How many episodes does Game of Thrones have?". Unlike the looping strategy, which iteratively retrieves information across multiple instances (e.g., years for awardWonBy), the direct strategy generates the required object entity in a single step for each relation.
Our empirical findings show that providing additional context about the subject entity in the
prompt directs the LLM’s attention to relevant parts of the training data, thereby improving
their ability to retrieve accurate answers and essential information, such as the first year an
award was conferred. Obtaining this information can also help define where the loop starts for
the awardWonBy relation. Context-aware prompting involves enriching the LLM prompt with
relevant background information about the subject [
Central to our methodology is the implementation of persona-based prompting, a role-playing
strategy that enriches the contextual engagement of our system. Role-play prompting is a prompt
engineering technique that allows LLMs to adopt specific personas or characters, guiding their
responses to align with the expert knowledge, thereby enhancing LLMs’ ability to generate
more precise and factually correct responses, as evidenced by studies such as [
Re-prompting is a technique in prompt engineering where an LLM is asked the same question
again to improve the quality of its responses [
Entity disambiguation is the final step in identifying and mapping the predicted object entities to their correct references in Wikidata. We use a straightforward disambiguation function that returns the Wikidata ID of an item. We clean the entity strings by removing unwanted characters such as quotes and parentheses. Specifically, for the awardWonBy relation, we employ the Spacy library to remove titles and extract person names accurately, ensuring precise identification and mapping of award recipients. For the companyTradesAtStockExchange relation, we observed that the LLM would sometimes answer with the complete stock exchange name as well as its abbreviation. In such cases, we split the LLM response into two separate entities, one for the full name and one for the abbreviation. We then attempt to disambiguate using the full name and the abbreviation, depending on which matches the Wikidata entry. For example, if the LLM outputs "New York Stock Exchange (NYSE)," we would search Wikidata for both "New York Stock Exchange" and "NYSE" to find the correct Wikidata ID. This method enhances the accuracy of linking the mentioned stock exchange to its oficial record in Wikidata.
This section includes an overview and discussion of our final results as well as an in-depth analysis of our model setup.
The dataset used in ISWC 2024 LM-KBC Challenge [
+ chain-of-thought + context-aware prompting + re-prompting + role-play prompting
In our work, we utilize the Llama3-8B-Instruct as the foundation of our final model setup, which incorporates various enhancements to the baseline pipeline. The baseline pipeline is comprised of dual prompting as well as the looping and direct strategies (as detailed in sections 3.1, 3.2, 3.3). It alone shows solid results on the validation dataset, with a macro precision of 0.662, a macro recall of 0.454, and a macro F1-score of 0.438. However, this setup results in 173 empty predictions as opposed to the expected 97, indicating the model’s weakness in producing relevant outputs.
Incorporating chain-of-thought prompting into our baseline setup, where the model is asked
to provide explanations for its responses, leads to a slight improvement in performance metrics,
as shown in table 2. Requesting a justification of the answer from the model when generating
its response helps reduce instances of non-responses, as the model is encouraged to elaborate
on its reasoning process [
The inclusion of a re-prompting mechanism (section 3.6), where the model is prompted again if the initial response cannot be extracted, yields further improvements in the performance metrics. This demonstrates that iterative prompting helps in refining the model’s output, ensuring higher accuracy and fewer empty responses. A more robust extraction function could also lead to similar results. Finally, the use of role-play prompting with the inclusion of personas (section 3.5) further enhances the model’s ability to generate informed and contextually rich responses, guiding the model to engage with the data as if it were an expert in the relevant field.
Our experiments demonstrate the eficacy of combining a variety of prompting techniques. The integration of context-aware prompting, role-play prompting with personas, and dual prompting, along with the re-asking strategy, leads to significant improvements in the precision, recall, and overall F1 scores. These findings underscore the importance of providing LLMs with rich contextual cues and structured guidance to achieve high accuracy and detailed responses in knowledge probing.
The introduction of a persona into our prompt pipeline signicfiantly enhances the model’s ability to generate accurate and contextually relevant responses, particularly for entities it would typically be uncertain about or would claim to lack familiarity with as shown in table 3. We chose to examine the impact of diferent role-play instructions on the final results in order to determine the most suitable approach to our problem.
Simply using a general persona (e.g. a trivia show contestant) across all relations did not provide an improvement in our F1 scores. However, as can be seen in the dip in empty predictions this approach did still lead to the model attempting to generate answers to the set questions.
We determined that tailoring personas to each relation type (e.g. an expert financial analyst for the companyTradesAtStockExchange) provided a more specialized context, allowing the model to generate answers with greater precision and relevance. The greatest benefit, however, was observed when personas were tailored individually to each entity (e.g., the CFO of the specific company for the companyTradesAtStockExchange). By suggesting that the model knows everything there is to know about the specific entity, the instances of empty or unsure responses were significantly reduced. The scores reflect the enhanced ability of the model to provide expert-like responses, confirming the efectiveness of persona-based prompting in our model pipeline.
Choosing the right prompting strategy can significantly impact the final results, as shown in table 4. We initially found success using the looping strategy for the awardsWonBy relation, where the model was asked to identify the award winner for each year before aggregating the results. Encouraged by this improvement, we applied the same looping strategy to the seriesHasNumberOfEpisodes relation. We expected that this strategy would lead to more accurate results due to the LLM likely having encountered more information about the number of episodes per season rather than the total number of episodes. Additionally, when manually prompting the LLM for the total number of episodes, we observed that the answers would often contain the total number of episodes per season first, followed by an often incorrect summation, awardWonBy seriesHasNumberOfEpisodes
0.259 0.412
F1 0.009 which would be provided as the final answer. Our intuition followed that eliciting the number of episodes per season one by one and then summing up said results manually as part of the disambiguation would be more efective than directly asking for the total number of episodes.
However, as shown in table 4, the results did not align with our expectations. The direct strategy, which involves simply asking for the total number of episodes, significantly outperformed both looping strategies.
The poor performance of the looping strategies can be attributed to the additional complexity of our pipeline. As each season is processed individually before summing up all answers in the disambiguation function, there is an increased risk of errors due to the multiple cases in which the LLM can potentially hallucinate. In contrast, the direct strategy has a single point of error. By utilizing a single prompt, we reduce the likelihood of numerical hallucinations and misinterpretations.
Nevertheless, the results from the direct strategy are far from perfect, suggesting that we need to devise a better way of extracting numerical information. One potential approach would be to consolidate both strategies. By combining the simplicity of the direct strategy with the detailed step-by-step verification of the looping strategies, we might achieve a more accurate and reliable method for handling numerical data. This hybrid approach could mitigate the weaknesses of each strategy, leading to improved performance in extracting numerical information from LLMs.
Our final pipeline demonstrates significant improvements across multiple evaluation metrics on the test and validation datasets. Table 5 summarizes the macro average precision, recall, and F1-score for each relation, along with the overall averages. The results are split into two sets of columns: the first three columns represent the test set results, and the latter three represent the validation set results. Additionally, for zero-object cases, our pipeline achieves a precision of 0.757, recall of 0.791, and F1-score of 0.773. awardWonBy companyTradesAtStockExchange countryLandBordersCountry personHasCityOfDeath seriesHasNumberOfEpisodes Average 0.476 0.675 0.978 0.890 0.480 0.729
Recall 0.399 0.585 0.890 0.800 0.440 0.653
In the case of relations where null values are possible, such as countryLandBordersCountry, personHasCityOfDeath, and companyTradesAtStockExchange, the pipeline performs exceptionally well. The dual prompting method we employ ensures that instances where null values are present are correctly identified. This is also reflected in the high accuracy, precision, and F1 scores achieved for zero-object cases.
Despite the complexity of dealing with numeric objects, our pipeline achieves a moderate performance in seriesHasNumberOfEpisodes. The macro F1-scores of 0.440 on both test and validation sets indicate that while the model can handle numeric data, there is room for improvement. This suggests that additional refinements may be needed for relations involving numeric outputs.
Finally, the pipeline’s performance on the awardWonBy relation, which involves many objects per subject with multiple instances, was the lowest. On the test set, the model achieves a satisfactory performance given the complexity of the task. Despite the efectiveness of the looping strategy on the test set, the significant drop in recall and F1-score on the validation set indicates challenges in maintaining consistency across diferent datasets. This can be explained by diferences in the examples between the test and validation sets. Certain awards in the validation set such as "honorary doctor of Stockholm University" have proven to be a significant challenge to our pipeline however equivalent examples to them are not present in the test set. Moreover, the low-performance metrics on the awardWonBy relation are likely influenced by a small sample size. With such a limited number of examples, namely 10 per both test and validation split, the evaluation may not adequately reflect the model’s true performance capabilities. More examples are needed to provide an accurate assessment and to ensure that the looping strategy and other techniques are efectively enhancing the model’s performance for this relation.
We also noticed during our analysis that Wikidata does not always accurately reflect the ground truth, which poses a significant challenge for evaluating model performance. For instance, for the subject entity "Grammy Award For Best Rock Album," our model generates the names of the artists who have won the award. However, upon examining the expected answers, we found that sometimes the object entities reflect the name of the winning album, and other times they reflect the names of the artists. This inconsistency within the expected answers can lead to lower recall and F1 scores even when the model performs well at the given task.
The quality issue of Wikidata has been highlighted in previous works [
In the future, we plan to explore several enhancements to improve our methodology further.
One promising direction is the use of cloze-style prompting in our pipeline. Empirical evidence
has shown that cloze-style prompts, which frame the query as a fill-in-the-blank task, can lead
to great results on this task. Another area of focus will be the introduction of few-shot examples.
Our current results indicate that the LLM does not always format answers correctly, leading to
inaccuracies. By providing a few-shot learning setup, where the model is given several examples
of correctly formatted answers, we hope to enhance its ability to generate responses that adhere
to the desired format consistently. Additionally, we plan to experiment with diferent model
architectures. Our current work utilizes a decoder-only model, but in future research, we aim
to explore alternatives such as encoder-decoder models. These diferent architectures may
ofer unique advantages in processing complex relational tasks and generating more accurate,
contextually appropriate responses. Finally, handling numerical information more efectively is
crucial, as shown by the moderate performance in the seriesHasNumberOfEpisodes relation. To
that end, we intend to investigate various techniques for processing numeric data, inspired by
research on the numerical reasoning capabilities in language models [
In this work, we introduce a comprehensive approach that fuses various prompt engineering techniques to address KBC from pre-trained language models. Our primary contribution is using dual prompting and looping strategies to handle relations with possible null answers and more complex one-to-many relations efectively. Dual prompting systematically guides the model through a two-step questioning process, enhancing its ability to resolve ambiguities and leading to a reduction in hallucinations. The looping strategy, on the other hand, is designed to manage complex relations by breaking down tasks into smaller, manageable parts, ensuring accurate extraction of information across multiple instances.
In the context of the ISWC’24 KBC-LLM challenge, our methods lead to a significant improvement over the set baseline, achieving 0.653 and 0.629 Macro-F1 scores on the test and validation datasets respectively. The combination of context-aware prompting, role-play personas, and re-prompting mechanisms further refines the model’s outputs, reducing hallucinations and enhancing accuracy. These results underscore the potential of advanced prompt engineering techniques in leveraging pre-trained LLMs for efective knowledge extraction and emphasize the importance of tailored strategies to address specific relational challenges.
countryLandBordersCountry personHasCityOfDeath
awardWonBy companyTradesAtStockExchange
countryLandBordersCountry You are an enthusiastic and knowledgeable resident of {subject_entity}, deeply in love with your homeland, and always eager to share your wealth of knowledge about it. You take pride in your country’s history, culture, geography, and the intricate details of its borders. Your passion for {subject_entity} shines through in every conversation, and you have a talent for explaining why it’s such a wonderful place to live. You’re always ready to provide detailed and accurate information about the country’s neighboring countries and land borders, aiming to convince your friends and anyone who listens that {subject_entity} is the best place to call home. You vividly describe the landscapes, cities, and unique features of the country, highlighting its connections and relationships with its neighbors. Your love for {subject_entity} is infectious, and you hope to inspire others to appreciate it as much as you do. personHasCityOfDeath seriesHasNumberOfEpisodes awardWonBy You are an ardent admirer and devoted follower of {subject_entity}, whose life and achievements have profoundly influenced your own. Your admiration for {subject_entity} has led you to meticulously study every aspect of their biography, ensuring you stay updated with the most accurate and detailed information about their life. You know their story inside and out, from their early beginnings to their current status. Your dedication to {subject_entity} means you are well-versed in the significant events of their life, including the critical details of where they lived, worked, and if applicable, where they passed away. When someone inquires about {subject_entity}, you are not only eager but also exceptionally qualified to provide precise and comprehensive answers. Your deep respect and admiration drive you to share their legacy accurately, ensuring that others understand the importance and impact of {subject_entity} in the correct context.
You are the biggest fan of {subject_entity}. You have watched every episode multiple times and know all the details about the show’s seasons and episodes. Your love for the show drives you to stay updated with every bit of information, and you take pride in your deep knowledge of it. Your friends and family always come to you when they have questions about {subject_entity} because they know you have the answers. When it comes to the number of episodes per season and in total, you can recall this information efortlessly and accurately. You really want to have your close ones also watch the show. You believe that if you answer your friends and family’s questions correctly, they will start watching the show with you. Use your passion and expertise to provide detailed and precise answers about the show to your friends.
You are an aspiring recipient of the prestigious {subject_entity}, someone who has dedicated years to studying its history, past winners, and the significance of each accolade. Your passion for this award is unparalleled, and you know the details of its ceremonies, the recipients, and the criteria for winning by heart. Your knowledge is not just academic but deeply personal, as each fact and figure represents a step closer to your own dream. This drives you to provide accurate, thorough, and insightful answers regarding the award and its winners in all years since its inception. companyTradesAtStockExchange You are the Chief Financial Oficer (CFO) of {subject_entity}, a key executive responsible for managing the company’s financial actions. You possess an in-depth understanding of all financial matters related to the company, including detailed knowledge about the company’s stock, listing status, and financial performance. As the CFO, you are dedicated to truthfulness and transparency, ensuring that all stakeholders, including potential investors, have accurate and comprehensive information. You are highly knowledgeable about the stock exchange, regulatory requirements, and the company’s financial strategy. Your responses are characterized by precision, clarity, and reliability, as you aim to foster trust and confidence in {subject_entity}’s financial health and investment potential. countryLandBordersCountry personHasCityOfDeath
awardWonBy companyTradesAtStockExchange I’d love to know more about the ge- I’d love to know more about the geography from someone who truly ography from someone who truly understands and loves this coun- understands and loves this country. try. Does {subject_entity} share any Which countries share a land borland borders? I’ve been thinking of der with {subject_entity}? I’ve been moving here! Final answer should thinking of moving here! Final anbe yes or no. swer should be a list of countries. Your admiration for {subject_entity} Your admiration for {subject_entity} has undoubtedly led you to study has undoubtedly led you to study their life extensively. Has {sub- their life extensively. In which city ject_entity} died? I wish to learn did {subject_entity} die? I wish to more about such an influential fig- learn more about such an influenure. Final answer should be yes or tial figure. Final answer should be no. the city name.
dreamed of winning the {subject_entity}, you are well-versed in its history. Was the {subject_entity} awarded? Final answer should be yes or no.
must be aware of all pertinent details regarding the company’s stock.
stock exchange? As a potential investor, I would like to have more details before making a decision to invest. Final answer should be yes or no
must be aware of all pertinent details regarding the company’s stock.
No x <= current year ?
Yes Dual Prompt for year x : Contextual Information + Role-play Persona + Initial Prompt + Follow-up Prompt(s)
Llama-3-8B
[ Award Recipient(s) for year xi] [ All Award Recipient(s) ] x = x + 1 Llama-3-8B
Response If ‘Response’ not in desired format
No
Yes Re-prompting: Output: ObjectEntityID from the Llama-3-8B LLM. The process begins with an initial prompt to the LLM, followed by an evaluation of the response. If the response is not in the desired format, the model undergoes reprompting to adjust and improve the output. The process ends with entity disambiguation, ultimately producing the final output as the ObjectEntityID.
companyTradesAtStockExchange companyTradesAtStockExchange System Message: PERSONA: You are the Chief Financial Officer (CFO) of {subject_entity}, a key executive responsible for managing the company's financial actions… INSTRUCTION: If the question can be answered simply with a yes or no - response should only be [Yes] or [No] … User Message: ADDITIONAL INFORMATION: The legal form of the company {subject_entity} is {info}. Here is an extract from the Wikipedia page for {subject_entity}: {info} INITIAL PROMPT: As the CFO of {subject_entity} you must be aware of all pertinent details regarding the company's stock. Is the {subject_entity} listed on the stock exchange? As a potential investor, I would like to have more details before making a decision to invest. Final answer should be yes or no.
System Message: PERSONA: You are the Chief Financial Officer (CFO) of {subject_entity}, a key executive responsible for managing the company's financial actions… INSTRUCTION: If the question can be answered simply with a yes or no - response should only be [Yes] or [No] … User Message: ADDITIONAL INFORMATION: The legal form of the company {subject_entity} is {info}. Here is an extract from the Wikipedia page for {subject_entity}: {info} FOLLOW-UP PROMPT: As the CFO of {subject_entity} you must be aware of all pertinent details regarding the company's stock.
Where do shares of {subject_entity} trade? Final answer should consist of name(s) of stock exchange(s). (a) (b) the dual prompting strategy applied to elicit detailed responses from the persona of a Chief Financial
PERSONA: You are an aspiring recipient of the prestigious {subject_entity}, someone who has dedicated years to studying its history, past winners, and the significance of each accolade… INSTRUCTION: If the question can be answered simply with a yes or no response should only be [Yes] or [No] … PERSONA: You are an aspiring recipient of the prestigious {subject_entity}, someone who has dedicated years to studying its history, past winners, and the significance of each accolade… INSTRUCTION: If the question can be answered simply with a yes or no response should only be [Yes] or [No] … System Message: PROMPT: As the biggest fan of {subject_entity} you know everything there is to know about the show. How many episodes does {subject_entity} have? Final answer should be a number …