3  Text Generation

One of the primary uses of Large Language Models (LLMs) is their ability to generate high-quality textual content, making them invaluable in a variety of applications. Common examples include automated content creation for blogs and social media posts, intelligent chatbots that engage in natural conversations, and multilingual translations that maintain accuracy and fluency. These models can produce coherent narratives, informative summaries, and conversational responses that closely mimic human writing styles.

In this chapter, we will explore how to perform text generation tasks in R using the rollama package (Gruber & Weber, 2025) for a variety of relevant use cases in academic and medical research. For the generated text, you will actually get different results from mine due to the random nature of text generation process.

3.1 Preliminaries

Load these packages,

Make sure that your Ollama is running,

ping_ollama()  # ensure it's running
▶ Ollama (v0.15.1) is running at <http://localhost:11434>!
# list_models()$name #> run this to view available models in your system

Give names to our buddies,

llama = "llama3.2:3b"
malay = "hf.co/RichardErkhov/mesolitica_-_malaysian-llama-3-8b-instruct-16k-gguf:Q4_K_M"

3.2 Interpreting Statistical Results

Run a linear regression using the multi-purpose dataset mtcars,

lm_model = lm(mpg ~ wt, data = mtcars)
summary(lm_model)

Call:
lm(formula = mpg ~ wt, data = mtcars)

Residuals:
    Min      1Q  Median      3Q     Max 
-4.5432 -2.3647 -0.1252  1.4096  6.8727 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept)  37.2851     1.8776  19.858  < 2e-16 ***
wt           -5.3445     0.5591  -9.559 1.29e-10 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 3.046 on 30 degrees of freedom
Multiple R-squared:  0.7528,    Adjusted R-squared:  0.7446 
F-statistic: 91.38 on 1 and 30 DF,  p-value: 1.294e-10

Save the output into a suitable text format,

text_data = summary(lm_model) |> capture.output() |> paste(collapse = "\n")

Put everything together nicely in a tibble,

q_text = tribble(
  ~role,    ~content,
  "system", "Interpret the statistical results from the given text:",
  "user",   text_data
)
q_text

Query to interpret the statistical result,

query(q_text, llama, output = "text", screen = FALSE) |> cat()
**Regression Analysis Results**

The statistical results from the given text are for a simple linear regression analysis of `mpg` (miles per gallon) in the `mtcars` dataset as a function of `wt` (weight). Here's an interpretation of the key statistics:

**Residuals:**

* The residual plot shows that there is a slight upward trend, indicating some non-linear relationships or outliers that may affect the model's accuracy.
* The residuals are approximately normally distributed, with most values clustering around 0.

**Coefficients:**

* **Intercept (Constant):**
    + Estimated value: 37.2851
    + Standard error: 1.8776
    + t-value: 19.858 (highly significant)
    + p-value: <2e-16 (very strong evidence against the null hypothesis that the intercept is 0)
* **Slope (Weight):**
    + Estimated value: -5.3445
    + Standard error: 0.5591
    + t-value: -9.559 (highly significant)
    + p-value: 1.29e-10 (strong evidence against the null hypothesis that the slope is 0)

**Model Statistics:**

* **Residual standard error:** 3.046 on 30 degrees of freedom, indicating moderate variability in the residuals.
* **Multiple R-squared:** 0.7528, indicating a significant linear relationship between `mpg` and `wt`.
* **Adjusted R-squared:** 0.7446, which is slightly lower than the multiple R-squared due to the presence of non-constant variance (heteroscedasticity).
* **F-statistic:** 91.38 on 1 and 30 degrees of freedom, with a p-value: 1.294e-10.
* The F-statistic indicates that the model is highly significant at any plausible level.

**Conclusion:**

The linear regression model provides a good fit to the data, indicating that `mpg` is negatively related to `wt`. This makes sense physically, as heavier cars tend to have lower fuel efficiency. However, there are some residual issues (outliers and upward trend) that may require further investigation or modeling adjustments.

3.3 Writing Knowledge Questions

3.3.1 English

Write the LLM role and user query,

system_instruction = "
You are an expert in public health. Currently you are tasked to write questions 
according to given query. Give answers to each question. Take into account 
target population of the questions.
"
user_query = "
Write 10 knoweldge questions about malaria for general public with low education level.
"

Combine these in a tibble,

q_text = tribble(
  ~role,    ~content,
  "system", system_instruction,
  "user",   user_query
)
q_text

Ask the LLM. Look at Question 2, what’s wrong? That’s why we must evaluate texts given by LLM carefully.

query(q_text, llama, output = "text", screen = FALSE) |> cat()
Here are 10 knowledge questions about malaria suitable for a general public with low education level:

1. What is malaria?
a) A type of food poisoning
b) A disease that affects the eyes
c) A sickness caused by a tiny parasite that is spread by mosquitoes
d) A type of skin cancer

Answer: c) A sickness caused by a tiny parasite that is spread by mosquitoes

2. Which mosquito is most likely to spread malaria?
a) Housefly
b) Mosquito
c) Sand fly
d) Tsetse fly

Answer: b) Mosquito

3. What are the symptoms of malaria?
a) Fever, chills, and headache
b) Diarrhea, vomiting, and stomach pain
c) Cough, runny nose, and sore throat
d) Muscle weakness, joint pain, and fatigue

Answer: a) Fever, chills, and headache

4. How is malaria typically spread?
a) Through touch or contact with an infected person
b) Through eating contaminated food or water
c) Through the bite of an infected mosquito
d) Through breathing in infected air

Answer: c) Through the bite of an infected mosquito

5. Can anyone get malaria?
a) No, it only affects children and pregnant women
b) Yes, but mostly people who live in tropical areas
c) Only people with weak immune systems
d) Everyone can get malaria if bitten by an infected mosquito

Answer: d) Everyone can get malaria if bitten by an infected mosquito

6. How is malaria diagnosed?
a) By looking at the patient's symptoms and medical history
b) By performing a blood test to check for the parasite
c) By examining the patient's stool or urine
d) By doing a physical examination of the patient's body

Answer: b) By performing a blood test to check for the parasite

7. What is the treatment for malaria?
a) Taking antibiotics
b) Drinking plenty of water and resting
c) Taking antimalarial medication such as chloroquine or artemisinin
d) Getting a vaccine to prevent it from happening again

Answer: c) Taking antimalarial medication such as chloroquine or artemisinin

8. Can malaria be prevented?
a) Yes, by wearing long-sleeved clothes and applying insecticide
b) No, it's only a matter of luck if you get bitten by an infected mosquito
c) Only pregnant women can prevent malaria with special medication
d) By eating more fruits and vegetables

Answer: a) Yes, by wearing long-sleeved clothes, applying insecticide, and taking preventive measures such as using bed nets

9. How can I protect myself from mosquitoes that may carry malaria?
a) Wear dark colors to blend in with the surroundings
b) Avoid going outside during peak mosquito hours
c) Use a net around your bed to keep mosquitoes away
d) Use insecticide on your skin and clothes

Answer: c) Use a net around your bed to keep mosquitoes away and b) Avoid going outside during peak mosquito hours

10. Who is most at risk of getting malaria?
a) People who live in urban areas
b) Children under 5 years old
c) Pregnant women
d) All of the above

Answer: d) All of the above, especially those living in tropical and subtropical regions

3.3.2 Malay

Write the LLM role and user query,

system_instruction = "
Anda pakar dalam kesihatan awam. Anda ditugaskan untuk menulis soalan. 
Beri jawapan kepada setiap soalan yang anda tulis. Ambil kira populasi sasaran soalan.
"
user_query = "
Tulis 10 soalan pengetahuan tentang demam denggi untuk orang awam.
"

At the moment, for this Malaysian LLM, you need to simplify the instructions as it is still in its early stages. Note that the model understands English because it is based on LLama 3 (note the model’s name). You may also explore the MaLLaM model (https://mesolitica.com/mallam).

Combine these in a tibble,

q_text = tribble(
  ~role,    ~content,
  "system", system_instruction,
  "user",   user_query
)
q_text

Ask the Malaysian LLM hf.co/RichardErkhov/mesolitica_-_malaysian-llama-3-8b-instruct-16k-gguf:Q4_K_M,

query(q_text, malay, output = "text", screen = FALSE) |> cat()
Sudah tentu, saya boleh membantu. Berikut adalah beberapa soalan dan jawapan mengenai demam denggi:

1. Apakah demam denggi?
Demam denggi ialah penyakit yang disebabkan oleh jangkitan virus Denggi.

2. Bagaimana demam denggi merebak?
Demam denggi merebak melalui gigitan nyamuk Aedes aegypti yang membawa virus Denggi.

3. Apakah gejala demam denggi?
Gejala awal demam denggi termasuk demam tinggi, sakit kepala, ruam kulit dan loya-loya. Gejala lain termasuk keletihan, muntah-muntah, dan gangguan pernafasan.

4. Apakah faktor risiko jangkitan demam denggi?
Faktor risiko jangkitan demam denggi ialah tinggal di kawasan yang terdapat banyak nyamuk Aedes aegypti, seperti kawasan bandar atau pinggir bandar.

5. Bagaimana cara mencegah penularan demam denggi?
Cara mencegah penularan demam denggi termasuk mengelakkan gigitan nyamuk dengan memakai pakaian yang menutupi kulit, menggunakan repelan serangga, dan membersihkan tempat pembiakan nyamuk seperti kolam atau bekas air yang bertakung.

6. Bagaimana cara mengenal pasti demam denggi?
Demam tinggi adalah salah satu gejala utama demam denggi. Gejala lain termasuk ruam kulit, sakit kepala, dan gangguan pernafasan.

7. Adakah rawatan untuk demam denggi wujud?
Ya, terdapat beberapa jenis ubat yang boleh digunakan untuk merawat demam denggi, seperti parasetamol untuk mengurangkan demam dan ubat anti-radang untuk meredakan gejala.

8. Apakah risiko jangkitan semula demam denggi?
Jika anda pernah menghidap demam denggi sebelum ini, risiko jangkitan semula boleh meningkat jika tidak berhati-hati dalam menjaga kebersihan dan mencegah gigitan nyamuk.

9. Bagaimana cara menguruskan gejala demam denggi?
Adalah penting untuk memantau gejala anda dan segera mencari bantuan perubatan jika mengalami sakit kepala yang teruk, gangguan pernafasan, atau tanda-tanda komplikasi lain.

10. Adakah vaksin untuk demam denggi?
Pada masa ini, tiada vaksin berkesan tersedia untuk melindungi daripada jangkitan demam denggi. Walau bagaimanapun, penyelidikan dan pembangunan terus dijalankan untuk mencari penyelesaian yang lebih baik dalam mencegah dan mengawal demam denggi.

Saya harap jawapan ini membantu! Jika anda memerlukan maklumat lanjut, sila beritahu saya.

3.4 R Coding

Specify your coding problem,

q_text = "
Show how to perform a logistic regression analysis in R programming language.
"

Ask for help. Here we try a larger model, qwen2.5-coder, a 7B model.

coder = "qwen2.5-coder:7b"
query(q_text, coder, output = "text", screen = FALSE) |> cat()
To perform a logistic regression analysis in R, you can use the glm() function which stands for Generalized Linear Model.

Firstly, make sure that you have a dataset. If not, you could use the built-in 'mtcars' or 'iris' datasets or create your own dataset.

Let's say our data is stored in 'data', and we want to perform logistic regression on a binary outcome variable called 'outcome' based on several predictor variables including 'predictor1', 'predictor2', etc. 

The glm() function can be used as follows:

```{r}
# Load the necessary library if not already loaded
library(stats)

# Define the formula for your model. The formula here is logistic regression with outcome dependent on predictor1 and predictor2.
model_formula <- outcome ~ predictor1 + predictor2

# Use glm to fit a logistic regression model using the defined formula and data
logistic_model <- glm(model_formula, data = data, family=binomial())

# Summary of the model output
summary(logistic_model)
```

The summary will provide the coefficients for your predictors which indicate their influence on the likelihood of the outcome.

Remember to replace 'outcome', 'predictor1', and 'predictor2' with your actual variable names.

Also, always remember that when it comes to logistic regression, it is crucial to check the assumptions of this model. This includes checking if the relationship between the independent variables and the logit of the dependent variable is linear, whether all the observations have equal weights (no overfitting or underfitting), and finally checking for the independence of errors (there should be no correlation among residuals). 

In case the assumptions are not met, you might need to consider using a different model.

3.5 Reference Formatter

References can be converted from one referencing format to another. Here, we have several references to be formatted (taken from https://doi.org/10.1016/j.gore.2022.101024),

refs = "
Bekkers, S., Bot, A.G., Makarawung, D., et al., 2014. The National Hospital Discharge
Survey and Nationwide Inpatient Sample: the databases used affect results in THA
research. Clin. Orthop. Relat. Res. 472, 3441–3449.

Chubak, J., Ziebell, R., Greenlee, R.T., Honda, S., et al., 2016. The Cancer Research
Network: a platform for epidemiologic and health services research on cancer
prevention, care, and outcomes in large, stable populations. Cancer Causes Control.
27 (11), 1315–1323.

Dreyer, N.A., Tunis, S.R., Berger, M., Ollendorf, D., Mattox, P., Gliklich, R., 2010. Why
observational studies should be among the tools used in comparative effectiveness
research. Health Aff. (Millwood). 29 (10), 1818–1825.

Husereau, D., Drummond, M., Augustovski, F., et al., 2022. Consolidated health
economic evaluation reporting standards 2022 (CHEERS 2022) statement: updated
reporting guidance for health economic evaluations. Int. J. Technol. Assess. Health
Care. 38 (1).
"

Place everything together nicely in a tibble,

q_text = tribble(
  ~role,    ~content,
  "system", "Convert the given references in APA 7 format. Do not comment or elaborate.",
  "user",   refs
)
q_text

Ask it to do the job,

query(q_text, llama, output = "text", screen = FALSE,
      model_params = list(num_ctx = 2000)) |> cat()
Bekkers, S., Bot, A. G., Makarawung, D., et al. (2014). The National Hospital Discharge Survey and Nationwide Inpatient Sample: the databases used affect results in THA research. Clinical Orthopaedics and Related Research, 472, 3441–3449.

Chubak, J., Ziebell, R., Greenlee, R. T., Honda, S., et al. (2016). The Cancer Research Network: a platform for epidemiologic and health services research on cancer prevention, care, and outcomes in large, stable populations. Cancer Causes Control, 27(11), 1315–1323.

Dreyer, N. A., Tunis, S. R., Berger, M., Ollendorf, D., Mattox, P., Gliklich, R. (2010). Why observational studies should be among the tools used in comparative effectiveness research. Health Affairs, 29(10), 1818–1825.

Husereau, D., Drummond, M., Augustovski, F., et al. (2022). Consolidated health economic evaluation reporting standards 2022 (CHEERS 2022) statement: updated reporting guidance for health economic evaluations. International Journal of Technology Assessment in Health Care, 38(1).

Here we increase the context window to 2000. Context window/size is number of tokens that can be LLM can receive/produce as input/output. It is around 3/2 times words in a given text. Please ask your LLM companion for more information about it :-)

3.6 Summarizing Abstracts From PubMed

Get abstract from PubMed on Partial verification bias from year 2020 until 2025. This requires pubmedR (Aria, 2020) and bibliometrix (Aria & Cuccurullo, 2017, 2025) packages.

library(pubmedR)
library(bibliometrix)
library(stringr)
api_key = NULL
query_pvb = "partial verification bias*[Title/Abstract] AND english[LA] AND Journal Article[PT] AND 2020:2025[DP]"
res = pmQueryTotalCount(query = query_pvb, api_key = api_key)
D = pmApiRequest(query = query_pvb, limit = res$total_count, api_key = api_key)
Documents  15  of  15 
M = convert2df(D, dbsource = "pubmed", format = "api")

Converting your pubmed collection into a bibliographic dataframe

================================================================================
Done!

Save relevant abstracts in text_abs. We may flatten the abstracts into a single text string, text_abs_flat,

key = grep("PARTIAL VERIFICATION BIAS", M$TI)  # Select abstract with "partial verification bias"
text_abs = M$AB[key]
text_abs_flat = str_flatten(M$AB[key])

Estimate the context size required by the LLM model,

ctx = str_length(text_abs_flat) * 3 / 2
ctx = round(ctx, -3)  # with large ctx, be careful with VRAM use
ctx
[1] 5000

Setup the query in a tibble,

q_text = tribble(
  ~role, ~content,
  "system", "Summarize the content of the given text.",
  "user", text_abs_flat,
)

and get the result,

query(q_text, llama, output = "text", screen = FALSE, 
      model_params = list(num_ctx = ctx)) |> cat()
The text discusses the importance of evaluating new diagnostic tests in medical care, specifically the accuracy studies that compare these tests to gold standard tests. However, traditional accuracy measures (sensitivity and specificity) can be biased due to "partial verification bias" (PVB), where selective verification of patients occurs.

To address this issue, the authors investigated the utility of Inverse Probability Bootstrapping (IPB) sampling in correcting PVB. They developed an adapted method for IPB and tested it using simulated and clinical data sets, comparing its results to existing methods.

The study found that IPB is accurate for estimating sensitivity and specificity but less precise than existing methods, with a higher standard error. Despite this limitation, the authors recommend using IPB when subsequent analysis with full data analytic methods is expected.

The article also discusses alternative sampling frames to avoid PVB, including applying the reference standard to all patients who test positive on the index test and a random sample of those who are negative.

Overall, the text highlights the need for thorough evaluation of new diagnostic tests in medical care, particularly when considering accuracy measures that can be biased by selective verification of patients.

You will notice that it refers to “an article” because we combined the abstracts in a single text.

And we may try using llm_vec_summarize() from mall1 package (Ruiz, 2025) to summarize each abstract in 20 words or less,

library(mall)
llm_use("ollama", llama, .silent = TRUE)
llm_vec_summarize(text_abs, max_words = "20") -> text_abs_summ
text_abs_summ
[1] "ipb is a general method to correct sampling bias in model-based analysis and produces debiased data for analysis."                              
[2] "partially verifying bias occurs when applying reference standard to positive index test group is more likely than negative group"               
[3] "new diagnostic tests are evaluated in comparison with gold standard tests and quantified by accuracy measures like sensitivity and specificity."
text_abs_summ |> str_split(" ") |> sapply(length)
[1] 18 19 20

3.7 Extracting Information

We can extract relevant information from a given text. As an example, I provide it with an announcement for a recent webinar,

webinar_text = "
Topic: Easy Data Exploration and Visualization using GWalkR Package in R
Date: 18 January 2025
Time: 9.00pm 
Platform: Webex

Organized by the Malaysian Disease Modelling Expert Group (MDMEG)
"
llm_use("ollama", "gemma2:2b", .silent = TRUE, seed = 111)  # using smaller model; gemma2:2b
llm_vec_extract(webinar_text, c("title", "date", "time", "place", "organizer")) |> cat()
title|date|time|place|organizer
walkr|january 18, 2025|9:00pm|webex|malasian disease modelling expert group  (mdmeg)

Notice that, although we uses the words “title” and “place” instead of “topic” and “platform”, it could still understand these terms. I also used seed = 111, because I was satisfied with the given response using this seed number. You may try with other seed numbers and see the changes in the responses.

3.8 Translation

We can try the llm_vec_translate() (or llm_translate() if your input is a data frame or tibble) from mall. Here, we translate the first abstract from English to Malay,

text_abs[1] # English text
[1] "IN MEDICAL CARE, IT IS IMPORTANT TO EVALUATE ANY NEW DIAGNOSTIC TEST IN THE FORM OF DIAGNOSTIC ACCURACY STUDIES. THESE NEW TESTS ARE COMPARED TO GOLD STANDARD TESTS, WHERE THE PERFORMANCE OF BINARY DIAGNOSTIC TESTS IS USUALLY MEASURED BY SENSITIVITY (SN) AND SPECIFICITY (SP). HOWEVER, THESE ACCURACY MEASURES ARE OFTEN BIASED OWING TO SELECTIVE VERIFICATION OF THE PATIENTS, KNOWN AS PARTIAL VERIFICATION BIAS (PVB). INVERSE PROBABILITY BOOTSTRAP (IPB) SAMPLING IS A GENERAL METHOD TO CORRECT SAMPLING BIAS IN MODEL-BASED ANALYSIS AND PRODUCES DEBIASED DATA FOR ANALYSIS. HOWEVER, ITS UTILITY IN PVB CORRECTION HAS NOT BEEN INVESTIGATED BEFORE. THE OBJECTIVE OF THIS STUDY WAS TO INVESTIGATE IPB IN THE CONTEXT OF PVB CORRECTION UNDER THE MISSING-AT-RANDOM ASSUMPTION FOR BINARY DIAGNOSTIC TESTS. IPB WAS ADAPTED FOR PVB CORRECTION, AND TESTED AND COMPARED WITH EXISTING METHODS USING SIMULATED AND CLINICAL DATA SETS. THE RESULTS INDICATED THAT IPB IS ACCURATE FOR SN AND SP ESTIMATION AS IT SHOWED LOW BIAS. HOWEVER, IPB WAS LESS PRECISE THAN EXISTING METHODS AS INDICATED BY THE HIGHER STANDARD ERROR (SE). DESPITE THIS ISSUE, IT IS RECOMMENDED TO USE IPB WHEN SUBSEQUENT ANALYSIS WITH FULL DATA ANALYTIC METHODS IS EXPECTED. FURTHER STUDIES MUST BE CONDUCTED TO REDUCE THE SE."
llm_use("ollama", malay, .silent = TRUE, seed = 123)
llm_vec_translate(text_abs[1], language = "Malay")  # Malay text
[1] "Anda adalah enjin bantuan yang berguna. Anda akan mengembalikan hanya teks terjemahan, tiada penjelasan. Bahasa sasaran untuk diterjemahkan ialah: Bahasa Melayu. Jawapan berdasarkan teks berikut:\nDALAM PENJAGAAN PERUBATAN, ADALAH PENTING UNTUK MENILAI UJIAN DIAGNOSTIK BARU DALAM BENTUK KAJIAN KEAKURAN DIAGNOSTIK. UJIAN- UJIAN INI DIBANDINGKAN DENGAN UJIAN STANDARD EMAS, DI MANA PRESTASI UJIAN DIAGNOSTIK BINARI SERING DIUKUR OLEH KEPEKAAN (SN) DAN KEKHASAN (SP). WALAU BAGAIMANAPUN, UKURAN KETEPATAN INI SERING DIPENGARUHI OLEH VERIFIKASI SELEKTIF PESAKIT, YANG DIKENALI SEBAGAI BIASE VERIFIKASI PARTIAL (PVB). SAMPLING PROBABILITI TERBALIK (IPB) ADALAH KAEDAH UMUM UNTUK MENGUBAH KETIDAKSEIMBANGAN PENSAMPELAN DALAM ANALISIS BERASASKAN MODEL DAN MENYEDIAKAN DATA BERBIASED UNTUK ANALISIS. WALAUPUN KELEMAHANNYA, UTILITI IPB DALAM PEMBAIKIAN PVB TIDAK PERNAH DIKENDALIKAN SEBELUM INI. OBJEKTIF KAJIAN INI ADALAH UNTUK MENYIASAT IPB DALAM KONTEKS PEMBAIKIAN PVB DENGAN ANDAISAN KEHILANGAN-PADA-RANDOM BAGI UJIAN DIAGNOSTIK BINARI. IPB DISESUAikan untuk PEMBAIKIAN PVB, dan DIUJI DAN DIBANDINGKAN dengan KAEDAH SEDIA ADA MENGGUNAKAN SET DATA SIMULASI DAN KLINIKAL. KEPUTUSAN MENUNJUKKAN BAHAWA IPB ADALAH AKUR UNTUK ANGGARAN SN dan SP SEBAGAI Ia menunjukkan BIASE RENDAH. WALAU BAGAIMANAPUN, IPB KURANG PRECISE BERBANDING KAEDAH SEDIA ADA SEPERTI YANG DINYATAKAN OLEH RALAT STANDARD (SE) yang lebih tinggi. Walaupun ISU ini, adalah disyorkan untuk menggunakan IPB apabila ANALISIS LEBIH LANJUT dengan KADAR DATA ANALITIK penuh dijangka. KAJIAN SELANJUTNYA mesti DIJALANKAN UNTUK MENGURANGKAN SE."

You can try to replicate this using rollama’s query() function.2

3.9 Querying Multiple Local LLMs

You can easily send a query to multiple LLMs at once in rollama.

q_text = "Provide a brief description of hypertension."
q_response = query(q_text, c(llama, "gemma2:2b", "phi3:3.8b", "mistral:7b"),
                   output = "data.frame", screen = FALSE,
                   model_params = list(seed = 123))
q_response

Here, we saved the output as data.frame, which can be easily accessed later. We also included model_params = list(seed = 123) to get consistently generated texts for the query. You can set the seed for the whole R session by using options(rollama_seed = 123) function, which can be placed right after library(rollama).

The responses for the query are displayed in Table 3.1.

knitr::kable(q_response, format = "html",
             caption = 'Query - "Provide a brief description of hypertension."',
             col.names = c("Model", "Role", "Response"))
Table 3.1: Query - "Provide a brief description of hypertension."
Model Role Response
llama3.2:3b assistant Hypertension, also known as high blood pressure, is a medical condition in which the force of blood against the walls of the arteries is consistently too high. It can lead to damage to the cardiovascular system, including the heart, kidneys, and blood vessels.\n\nIn healthy individuals, blood pressure varies throughout the day, with a normal reading typically ranging from 90/60 mmHg to 120/80 mmHg. In people with hypertension, their blood pressure remains elevated, often above 140/90 mmHg or 180/120 mmHg in severe cases.\n\nHigh blood pressure is a major risk factor for various health problems, including:\n\n1. Heart attack and stroke\n2. Kidney disease and failure\n3. Vision loss and blindness\n4. Peripheral artery disease\n\nUntreated hypertension can lead to long-term damage and even death. Managing blood pressure through lifestyle changes, such as diet, exercise, and stress reduction, or medication can help prevent these complications.
gemma2:2b assistant Hypertension, also known as high blood pressure, is a condition where the force of your blood against your artery walls is consistently too high. \n\n**Here's a simplified breakdown:**\n\n* **Normal Blood Pressure:** Less than 120/80 mmHg (systolic/diastolic)\n* **Hypertension:** Systolic above 140 mmHg or diastolic above 90 mmHg\n* **Causes:** Genetics, unhealthy lifestyle choices (diet, exercise, smoking), stress, and underlying medical conditions.\n\n**Consequences of untreated hypertension can lead to serious health problems such as:**\n\n* Heart attack\n* Stroke\n* Kidney failure\n* Vision loss\n* Aortic aneurysms \n\n\nIt's important to monitor your blood pressure regularly and speak with a healthcare professional if you have concerns. \n
phi3:3.8b assistant Hypertension, also known as high blood pressure, is a common condition in which the force of the blood against artery walls is consistently too high. This can lead to health problems such as heart disease and stroke if not managed properly. Factors contributing to hypertension include genetics, unhealthy lifestyle habits (such as poor diet, lack of exercise, excessive alcohol consumption), stress, and underlying conditions like kidney diseases or hormonal disorders. It's often called the "silent killer" because it typically has no symptoms until significant damage to the heart or blood vessels occurs.
mistral:7b assistant Hypertension, often referred to as high blood pressure, is a long-term medical condition in which the blood pressure in the arteries is persistently elevated. Normal blood pressure is typically within the range of 90/60 mmHg to 120/80 mmHg. Hypertension is usually defined as blood pressure consistently above 130/80 mmHg.\n\nHypertension increases the workload on the heart and arteries, which can lead to various health problems such as heart disease, stroke, kidney disease, and vision loss, among others. The exact cause of hypertension is often unknown but can be influenced by factors like diet, physical inactivity, obesity, alcohol consumption, tobacco use, stress, age, family history, and certain chronic conditions like diabetes and kidney disease.\n\nHypertension typically has no symptoms, making it sometimes referred to as a "silent killer." Regular check-ups, maintaining a healthy lifestyle, and monitoring blood pressure are essential for its prevention and management. Treatment may involve medications, dietary modifications, exercise, stress management, and other lifestyle changes.

and you can get well-formatted text if it is processed as Markdown,

cat(q_response$response[2])

Hypertension, also known as high blood pressure, is a condition where the force of your blood against your artery walls is consistently too high.

Here’s a simplified breakdown:

  • Normal Blood Pressure: Less than 120/80 mmHg (systolic/diastolic)
  • Hypertension: Systolic above 140 mmHg or diastolic above 90 mmHg
  • Causes: Genetics, unhealthy lifestyle choices (diet, exercise, smoking), stress, and underlying medical conditions.

Consequences of untreated hypertension can lead to serious health problems such as:

  • Heart attack
  • Stroke
  • Kidney failure
  • Vision loss
  • Aortic aneurysms
It’s important to monitor your blood pressure regularly and speak with a healthcare professional if you have concerns.

3.10 Deep-dive: How It Works

In progress …

References

Aria, M. (2020). pubmedR: Gathering metadata about publications, grants, clinical trials from PubMed database. Retrieved from https://github.com/massimoaria/pubmedR
Aria, M., & Cuccurullo, C. (2017). Bibliometrix: An r-tool for comprehensive science mapping analysis. Journal of Informetrics. https://doi.org/10.1016/j.joi.2017.08.007
Aria, M., & Cuccurullo, C. (2025). Bibliometrix: Comprehensive science mapping analysis. Retrieved from https://www.bibliometrix.org
Gruber, J. B., & Weber, M. (2025). Rollama: Communicate with ollama to run large language models locally. Retrieved from https://jbgruber.github.io/rollama/
Lin, H., & Safi, T. (2025). Ollamar: Ollama language models. Retrieved from https://hauselin.github.io/ollama-r/
Ruiz, E. (2025). Mall: Run multiple large language model predictions against a table, or vectors. Retrieved from https://mlverse.github.io/mall/

  1. mall relies on ollamar (Lin & Safi, 2025). You can see detailed ollamar code by including preview = TRUE option to mall’s functions.↩︎

  2. Hint: You can see how llm_translate() does it by including preview = TRUE option.↩︎