Merge pull request #18 from edgararuiz/updates

Llama 3.2 integration

NAMESPACE

@@ -9,6 +9,7 @@ S3method(llm_sentiment,data.frame)
 S3method(llm_summarize,"tbl_Spark SQL")
 S3method(llm_summarize,data.frame)
 S3method(llm_translate,data.frame)
+S3method(m_backend_prompt,mall_llama3.2)
 S3method(m_backend_prompt,mall_session)
 S3method(m_backend_submit,mall_ollama)
 S3method(m_backend_submit,mall_simulate_llm)

R/llm-classify.R

@@ -24,7 +24,7 @@
 #'
 #' data("reviews")
 #'
-#' llm_use("ollama", "llama3.1", seed = 100, .silent = TRUE)
+#' llm_use("ollama", "llama3.2", seed = 100, .silent = TRUE)
 #'
 #' llm_classify(reviews, review, c("appliance", "computer"))
 #'

R/llm-custom.R

@@ -15,7 +15,7 @@
 #'
 #' data("reviews")
 #'
-#' llm_use("ollama", "llama3.1", seed = 100, .silent = TRUE)
+#' llm_use("ollama", "llama3.2", seed = 100, .silent = TRUE)
 #'
 #' my_prompt <- paste(
 #'   "Answer a question.",

R/llm-extract.R

@@ -17,7 +17,7 @@
 #'
 #' data("reviews")
 #'
-#' llm_use("ollama", "llama3.1", seed = 100, .silent = TRUE)
+#' llm_use("ollama", "llama3.2", seed = 100, .silent = TRUE)
 #'
 #' # Use 'labels' to let the function know what to extract
 #' llm_extract(reviews, review, labels = "product")

R/llm-sentiment.R

@@ -16,7 +16,7 @@
 #'
 #' data("reviews")
 #'
-#' llm_use("ollama", "llama3.1", seed = 100, .silent = TRUE)
+#' llm_use("ollama", "llama3.2", seed = 100, .silent = TRUE)
 #'
 #' llm_sentiment(reviews, review)
 #'

R/llm-summarize.R

@@ -12,7 +12,7 @@
 #'
 #' data("reviews")
 #'
-#' llm_use("ollama", "llama3.1", seed = 100, .silent = TRUE)
+#' llm_use("ollama", "llama3.2", seed = 100, .silent = TRUE)
 #'
 #' # Use max_words to set the maximum number of words to use for the summary
 #' llm_summarize(reviews, review, max_words = 5)

R/llm-translate.R

@@ -12,7 +12,7 @@
 #'
 #' data("reviews")
 #'
-#' llm_use("ollama", "llama3.1", seed = 100, .silent = TRUE)
+#' llm_use("ollama", "llama3.2", seed = 100, .silent = TRUE)
 #'
 #' # Pass the desired language to translate to
 #' llm_translate(reviews, review, "spanish")

R/llm-use.R

@@ -19,11 +19,11 @@
 #' \dontrun{
 #' library(mall)
 #'
-#' llm_use("ollama", "llama3.1")
+#' llm_use("ollama", "llama3.2")
 #'
 #' # Additional arguments will be passed 'as-is' to the
 #' # downstream R function in this example, to ollama::chat()
-#' llm_use("ollama", "llama3.1", seed = 100, temp = 0.1)
+#' llm_use("ollama", "llama3.2", seed = 100, temp = 0.1)
 #'
 #' # During the R session, you can change any argument
 #' # individually and it will retain all of previous

R/m-backend-prompt.R

@@ -4,6 +4,55 @@ m_backend_prompt <- function(backend, additional) {
   UseMethod("m_backend_prompt")
 }
 
+#' @export
+m_backend_prompt.mall_llama3.2 <- function(backend, additional = "") {
+  base_method <- NextMethod()
+  base_method$extract <- function(labels) {
+    no_labels <- length(labels)
+    col_labels <- paste0(labels, collapse = ", ")
+    plural <- ifelse(no_labels > 1, "s", "")
+    text_multi <- ifelse(
+      no_labels > 1,
+      "Return the response exclusively in a pipe separated list, and no headers. ",
+      ""
+    )
+    list(
+      list(
+        role = "user",
+        content = glue(paste(
+          "You are a helpful text extraction engine.",
+          "Extract the {col_labels} being referred to on the text.",
+          "I expect {no_labels} item{plural} exactly.",
+          "No capitalization. No explanations.",
+          "{text_multi}",
+          "{additional}",
+          "The answer is based on the following text:\n{{x}}"
+        ))
+      )
+    )
+  }
+  base_method$classify <- function(labels) {
+    labels <- process_labels(
+      x = labels,
+      if_character = "Determine if the text refers to one of the following: {x}",
+      if_formula = "If it classifies as {f_lhs(x)} then return {f_rhs(x)}"
+    )
+    list(
+      list(
+        role = "user",
+        content = glue(paste(
+          "You are a helpful classification engine.",
+          "{labels}.",
+          "No capitalization. No explanations.",
+          "{additional}",
+          "The answer is based on the following text:\n{{x}}"
+        ))
+      )
+    )
+  }
+  base_method
+}
+
 #' @export
 m_backend_prompt.mall_session <- function(backend, additional = "") {
   list(

R/m-defaults.R

@@ -5,8 +5,16 @@ m_defaults_set <- function(...) {
     nm <- names(new_args[i])
     defaults[[nm]] <- new_args[[i]]
   }
+  model <- defaults[["model"]]
+  split_model <- strsplit(model, "\\:")[[1]]
+  if (length(split_model) > 1) {
+    sub_model <- split_model[[1]]
+  } else {
+    sub_model <- NULL
+  }
   obj_class <- clean_names(c(
-    defaults[["model"]],
+    model,
+    sub_model,
     defaults[["backend"]],
     "session"
   ))

README.Rmd

@@ -16,7 +16,7 @@ library(dbplyr)
 library(tictoc)
 library(DBI)
 source("utils/knitr-print.R")
-mall::llm_use("ollama", "llama3.1", seed = 100, .cache = "_readme_cache")
+mall::llm_use("ollama", "llama3.2", seed = 100, .cache = "_readme_cache")
 ```
 
 # mall
@@ -78,9 +78,9 @@ This saves the data scientist the need to write and tune an NLP model.
 [Installation guide](https://hauselin.github.io/ollama-r/#installation)
 
 - Download an LLM model. For example, I have been developing this package using
-Llama 3.1 to test. To get that model you can run: 
+Llama 3.2 to test. To get that model you can run: 
   ```r
-  ollamar::pull("llama3.1")
+  ollamar::pull("llama3.2")
   ```
 
 ### With Databricks
@@ -218,10 +218,10 @@ use.
 Calling `llm_use()` directly will let you specify the model and backend to use.
 You can also setup additional arguments that will be passed down to the 
 function that actually runs the prediction. In the case of Ollama, that function
-is [`generate()`](https://hauselin.github.io/ollama-r/reference/generate.html). 
+is [`chat()`](https://hauselin.github.io/ollama-r/reference/chat.html). 
 
 ```{r, eval = FALSE}
-llm_use("ollama", "llama3.1", seed = 100, temperature = 0.2)
+llm_use("ollama", "llama3.2", seed = 100, temperature = 0)
 ```
 
 ## Key considerations
@@ -232,7 +232,8 @@ If using this method with an LLM locally available, the cost will be a long
 running time. Unless using a very specialized LLM, a given LLM is a general model. 
 It was fitted using a vast amount of data. So determining a response for each 
 row, takes longer than if using a manually created NLP model. The default model
-used in Ollama is Llama 3.1, which was fitted using 8B parameters. 
+used in Ollama is [Llama 3.2](https://ollama.com/library/llama3.2), 
+which was fitted using 3B parameters. 
 
 If using an external LLM service, the consideration will need to be for the 
 billing costs of using such service. Keep in mind that you will be sending a lot
@@ -260,34 +261,32 @@ library(classmap)
 
 data(data_bookReviews)
 
-book_reviews <- data_bookReviews |>
-  head(100) |>
-  as_tibble()
-
-glimpse(book_reviews)
+data_bookReviews |>
+  glimpse()
 ```
 As per the docs, `sentiment` is a factor indicating the sentiment of the review:
 negative (1) or positive (2)
 
 ```{r}
-length(strsplit(paste(book_reviews, collapse = " "), " ")[[1]])
+length(strsplit(paste(head(data_bookReviews$review, 100), collapse = " "), " ")[[1]])
 ```
 
 Just to get an idea of how much data we're processing, I'm using a very, very 
 simple word count. So we're analyzing a bit over 20 thousand words.
 
 ```{r}
-reviews_llm <- book_reviews |>
+reviews_llm <- data_bookReviews |>
+  head(100) |> 
   llm_sentiment(
     col = review,
-    options = c("positive", "negative"),
+    options = c("positive" ~ 2, "negative" ~ 1),
     pred_name = "predicted"
   )
 ```
 
-As far as **time**, on my Apple M3 machine, it took about 3 minutes to process,
-100 rows, containing 20 thousand words. Setting `temp` to 0.2 in `llm_use()`, 
-made the model run a bit faster.
+As far as **time**, on my Apple M3 machine, it took about 1.5 minutes to process,
+100 rows, containing 20 thousand words. Setting `temp` to 0 in `llm_use()`, 
+made the model run faster.
 
 The package uses `purrr` to send each prompt individually to the LLM. But, I did
 try a few different ways to speed up the process, unsuccessfully:
@@ -314,8 +313,8 @@ will not be of the "truth", but rather the package's results recorded in
 library(forcats)
 
 reviews_llm |>
-  mutate(fct_pred = as.factor(ifelse(predicted == "positive", 2, 1))) |>
-  yardstick::accuracy(sentiment, fct_pred)
+  mutate(predicted = as.factor(predicted)) |>
+  yardstick::accuracy(sentiment, predicted)
 ```
 
 ## Vector functions