Package 'conflibertR'

Description

ggplot2 version of plot.conflibert_al_session: the learning curve and the query-uncertainty trend across rounds.

Usage

## S3 method for class 'conflibert_al_session'
autoplot(object, which = c("all", "metrics", "uncertainty"), ...)

Arguments

Value

A ggplot object.

Description

Aggregated view of a conflibert_classify result: how many texts landed in each class, annotated with the average confidence.

Usage

## S3 method for class 'conflibert_classify'
autoplot(object, ...)

Arguments

Value

A ggplot object.

Description

Dot chart of every metric for every model in a conflibert_compare result, models ordered by their primary metric.

Usage

## S3 method for class 'conflibert_comparison'
autoplot(object, ...)

Arguments

Value

A ggplot object.

Description

Test-set confusion matrix heatmap for a conflibert_finetune result.

Usage

## S3 method for class 'conflibert_finetune'
autoplot(object, ...)

Arguments

Value

A ggplot object.

Description

Plots a conflibert_multilabel result. With several texts it aggregates: the share of texts flagged for each event category. With a single text it shows that text's category probabilities against the 0.5 decision threshold.

Usage

## S3 method for class 'conflibert_multilabel'
autoplot(object, ...)

Arguments

Value

A ggplot object.

Description

Aggregated view of a conflibert_ner result: type = "types" (default) counts entities per entity type; type = "entities" shows the most frequent individual entities, colored by type.

Usage

## S3 method for class 'conflibert_ner'
autoplot(object, type = c("types", "entities"), top_n = 12, ...)

Arguments

Value

A ggplot object.

Description

Opens a small Shiny gadget that shows each text in session$query alongside radio buttons for each class. Click Done to submit; the labels are returned as an integer vector ready to pass to conflibert_active_next.

Usage

conflibert_active_label(session, classes = NULL)

Arguments

Details

Requires the shiny and miniUI packages. In RStudio the gadget opens as a modal dialog; elsewhere it opens in your browser.

Value

An integer vector of labels (one per query row), or NULL if the user cancels.

Examples

data <- conflibert_example("active")
session <- conflibert_active_start(
  seed = data$seed, pool = data$pool, query_size = 5, epochs = 1
)
labels  <- conflibert_active_label(session)
session <- conflibert_active_next(session, labels)

Description

Takes a session from conflibert_active_start (or a previous call to this function), incorporates your labels for the current query, retrains the model on the full labeled set, and selects the next uncertain batch from the remaining pool.

Usage

conflibert_active_next(session, labels)

Arguments

Value

An updated conflibert_al_session. When the pool is exhausted, session$done is TRUE and session$query is empty.

Examples

data <- conflibert_example("active")
session <- conflibert_active_start(
  seed = data$seed, pool = data$pool, query_size = 5, epochs = 1
)
# label the queried texts (here using the bundled oracle labels)
labels  <- unname(data$pool_labels[session$query$text])
session <- conflibert_active_next(session, labels)
session$metrics

Description

Write the current session's model and tokenizer to disk. The result is a standard HuggingFace checkpoint that can be reloaded with any transformers tool.

Usage

conflibert_active_save(session, dir)

Arguments

Value

The directory path, invisibly.

Examples

data <- conflibert_example("active")
session <- conflibert_active_start(
  seed = data$seed, pool = data$pool, query_size = 5, epochs = 1
)
dir <- file.path(tempdir(), "my_al_model")
conflibert_active_save(session, dir)

Description

Train a classifier on a small labeled seed and select the most uncertain samples from an unlabeled pool for you to label. The returned session object tracks the labeled set, the pool, the current query, and metrics across rounds.

Usage

conflibert_active_start(
  seed,
  pool,
  dev = NULL,
  model = "ConfliBERT",
  task = c("binary", "multiclass"),
  strategy = c("entropy", "margin", "least_confidence"),
  diverse = FALSE,
  diversity_candidates = NULL,
  query_size = 10,
  epochs = 3,
  batch_size = 8,
  lr = 2e-05,
  max_seq_len = 512,
  use_lora = FALSE,
  lora_rank = 8,
  lora_alpha = 16,
  random_seed = 42
)

Arguments

Details

The typical workflow is:

1. Call conflibert_active_start() with your seed and pool.

2. Inspect session$query and assign labels.

3. Pass the labels to conflibert_active_next to retrain and query the next batch.

4. Repeat until the pool is exhausted or metrics plateau.

5. Persist with conflibert_active_save.

Value

An object of class "conflibert_al_session": a list with query (tibble of texts to label), metrics (tibble of metrics across rounds), round, labeled_n, pool_n, done, and internal state.

Examples

data <- conflibert_example("active")
session <- conflibert_active_start(
  seed = data$seed, pool = data$pool, query_size = 5, epochs = 1
)
session
session$query

Description

Returns TRUE when the "conflibert" Python environment exists and the core Python modules ('torch' and 'transformers') can be imported. It never installs anything; environment discovery uses filesystem checks only (it does not run the conda binary), and when no environment is found it returns FALSE without initializing Python, so it is cheap and safe to call on any system. It is used to guard the package's examples on machines without the backend (such as CRAN's check machines), and you can use it the same way in scripts that should degrade gracefully.

Usage

conflibert_available()

Details

The detection result is cached for the R session (if you have just run conflibert_install, restart R as its instructions say). Set the environment variable CONFLIBERTR_AVAILABLE to "true" or "false" to override the detection, e.g. to skip the backend-dependent examples during R CMD check on a machine that has the backend installed.

Value

TRUE or FALSE.

Examples

conflibert_available()

Description

Evaluate the pretrained ConfliBERT binary classifier against labeled data and compute accuracy, precision, recall, and F1.

Usage

conflibert_benchmark(texts, labels)

Arguments

Value

A tibble with one row and columns: accuracy, precision, recall, f1, n.

Examples

conflibert_benchmark(
  texts  = c("A bomb exploded.", "The weather was nice."),
  labels = c(1L, 0L)
)

Description

Classify text as conflict-related (Positive) or not (Negative) using the pretrained ConfliBERT binary classifier. Inference is batched, so long vectors are fast; a progress bar appears for large inputs.

Usage

conflibert_classify(text)

Arguments

Value

A tibble (with a themed print method) with columns:

text

The input text.

label

"Positive" or "Negative".

class

Integer (1 = positive, 0 = negative).

confidence

Probability of the predicted class.

prob_negative

Probability of the negative class.

prob_positive

Probability of the positive class.

Examples

conflibert_classify("A bomb exploded in the market.")

conflibert_classify(c(
  "Government troops clashed with rebels.",
  "The weather was sunny and warm."
))

Description

Fine-tune several base models on the same dataset and return a comparison table of test-set metrics. Useful for selecting the best architecture for your data.

Usage

conflibert_compare(
  train,
  dev,
  test,
  models = c("ConfliBERT", "BERT Base Uncased"),
  task = "binary",
  epochs = 3,
  batch_size = 8,
  lr = 2e-05,
  use_lora = FALSE,
  lora_rank = 8,
  lora_alpha = 16,
  seed = 42
)

Arguments

Value

A tibble with one row per model and columns for each metric plus runtime. It has a themed print() method that ranks models, and a plot() method comparing metrics.

Examples

data <- conflibert_example("binary")
comparison <- conflibert_compare(
  train  = data$train,
  dev    = data$dev,
  test   = data$test,
  models = c("ConfliBERT", "BERT Base Uncased"),
  task   = "binary",
  epochs = 1
)
comparison

Description

Load one of the bundled example datasets for testing fine-tuning and model comparison. Returns a list with train, dev, and test data frames ready to pass to conflibert_finetune or conflibert_compare.

Usage

conflibert_example(name = c("binary", "multiclass", "active"))

Arguments

Value

A named list of data frames (and a character vector for the active-learning pool). See name for the shape per dataset.

Examples

# Loading the bundled data is pure R and needs no Python backend:
data <- conflibert_example("binary")
data$train


# Fine-tune with example data (needs the Python backend)
data <- conflibert_example("binary")
result <- conflibert_finetune(
  train = data$train, dev = data$dev, test = data$test,
  model = "ConfliBERT", task = "binary", epochs = 1
)

# Compare models with example data
comparison <- conflibert_compare(
  train = data$train, dev = data$dev, test = data$test,
  models = c("ConfliBERT", "BERT Base Uncased"),
  task = "binary", epochs = 1
)

Description

Train a binary or multiclass text classifier on your own data using any of the supported base models (ConfliBERT, BERT, RoBERTa, ModernBERT, DeBERTa, DistilBERT).

Usage

conflibert_finetune(
  train,
  dev,
  test,
  model = "ConfliBERT",
  task = "binary",
  epochs = 3,
  batch_size = 8,
  lr = 2e-05,
  save_dir = NULL,
  use_lora = FALSE,
  lora_rank = 8,
  lora_alpha = 16,
  seed = 42
)

Arguments

Value

An object of class "conflibert_finetune" (a list, so all existing $ access keeps working) with:

metrics

Tibble of test-set metrics.

runtime

Training time in seconds.

predictions

Integer vector of predicted class labels.

probabilities

Matrix of class probabilities (rows = samples, columns = classes).

true_labels

Integer vector of true labels.

model_dir

Path where the model checkpoint was saved.

model, task

The base model name and task type.

It has a themed print() method and a plot() method showing the test-set confusion matrix.

Examples

train <- data.frame(
  text  = c("Troops advanced.", "Nice weather today."),
  label = c(1L, 0L)
)
result <- conflibert_finetune(train, dev = train, test = train, epochs = 1)
result$metrics

Description

Creates a Python environment and installs the packages needed to run ConfliBERT models (torch, transformers, and friends). Only needs to be run once.

Usage

conflibert_install(envname = "conflibert", method = "auto", qa = FALSE)

Arguments

Details

As of conflibertR 0.5.0 the backend is PyTorch-only: TensorFlow is no longer required, which makes installation considerably smaller and more reliable. If you have issues with the default virtualenv method, try method = "conda" instead.

Models downloaded from 'Hugging Face' are cached under $HF_HOME (default ~/.cache/huggingface) and the converted QA weights under ~/.cache/conflibertR (or $XDG_CACHE_HOME). Set those environment variables to relocate the caches.

Value

Invisible NULL. Called for its side effect.

Examples

## Not run: 
# Not run automatically anywhere: this installs software (it creates
# a Python environment and downloads several GB of dependencies), so
# it must only ever be run deliberately by the user.
conflibert_install()

# Include TensorFlow for the one-time QA weight conversion:
conflibert_install(qa = TRUE)

## End(Not run)

Description

Load a classifier saved by conflibert_finetune (when save_dir was specified) or conflibert_active_save. Returns a reusable classifier object you can pass to predict.

Usage

conflibert_load(dir)

Arguments

Value

An object of class "conflibert_classifier": a list with model, tokenizer, num_labels, and dir.

Examples

# Fine-tune on the bundled example data, saving the model...
data <- conflibert_example("binary")
dir <- file.path(tempdir(), "my_model")
conflibert_finetune(
  train = data$train, dev = data$dev, test = data$test,
  epochs = 1, save_dir = dir
)

# ...then reload it later and predict
clf <- conflibert_load(dir)
predict(clf, c("Troops advanced on the capital.", "Nice weather."))

Description

Returns the names of base models that can be used for fine-tuning and comparison.

Usage

conflibert_models()

Value

A character vector of model names.

Examples

conflibert_models()

Description

Score text against four event categories: Armed Assault, Bombing or Explosion, Kidnapping, and Other. Each category is scored independently. Inference is batched.

Usage

conflibert_multilabel(text)

Arguments

Value

A tibble (with themed print and plot methods) with columns:

doc_id

Integer index of the input text (only when length(text) > 1).

text

The input text.

label

Event category name.

probability

Score between 0 and 1.

predicted

Logical, TRUE if probability >= 0.5.

Examples

conflibert_multilabel("Insurgents kidnapped two aid workers near the border.")

Description

Identify persons, organizations, locations, weapons, and other entity types in text using the pretrained ConfliBERT NER model. Inference is batched, and printing highlights each entity in its source sentence.

Usage

conflibert_ner(text)

Arguments

Value

A tibble (with a themed print method) with columns:

doc_id

Integer. Which input text this entity came from (only present when length(text) > 1).

entity

Character. The entity text.

label

Character. Entity type (Person, Organisation, Location, Weapon, etc.).

score

Numeric. Mean model confidence for the entity span.

start, end

Integer. 1-based inclusive character offsets of the entity in the input text (use with substr()).

Examples

conflibert_ner("The soldiers attacked the village near Kabul.")

conflibert_ner(c(
  "NATO forces were deployed to the region.",
  "The UN Security Council met in New York."
))

Description

Extract an answer from a context passage given a question. Both arguments are vectorized: pass equal-length vectors to answer many questions in one call, or a single context with several questions (scalars are recycled).

Usage

conflibert_qa(context, question, details = FALSE)

Arguments

Details

The published question-answering checkpoint ships only 'TensorFlow' weights. The first call converts them to 'PyTorch' once and caches the result under ~/.cache/conflibertR (or $XDG_CACHE_HOME when set); subsequent calls reuse the cache and never touch 'TensorFlow' again. Downloaded 'Hugging Face' models are cached under $HF_HOME (default ~/.cache/huggingface). Both caches are written only on explicit, network-enabled calls, and can be relocated by setting those environment variables.

Value

With details = FALSE (default), a character vector of answers (a single string when one question is asked, exactly as in previous versions). With details = TRUE, a tibble with columns question, answer, score, start, end, and context.

Examples

conflibert_qa(
  context  = "The ceasefire was signed in Geneva on March 15th.",
  question = "Where was the ceasefire signed?"
)

# Several questions against one passage, with scores:
conflibert_qa(
  context  = "The ceasefire was signed in Geneva on March 15th.",
  question = c("Where was the ceasefire signed?",
               "When was the ceasefire signed?"),
  details  = TRUE
)

Description

Run a quick diagnostic of the Python backend: whether the "conflibert" environment exists, which Python is active, which required packages are importable, and what compute device will be used. Prints a checklist and gives specific advice when something is missing.

Usage

conflibert_status()

Value

Invisibly, a list with env_found, python, packages (named logical vector), device, and ok.

Examples

conflibert_status()

Description

Visualize how the model is improving across rounds. The default view is a two-panel plot: the learning curve on top (metrics vs training set size) and the uncertainty trend on the bottom (mean uncertainty of each round's queries). The uncertainty trend is a useful signal: when it flattens, the model is no longer finding informative samples.

Usage

## S3 method for class 'conflibert_al_session'
plot(x, which = c("all", "metrics", "uncertainty"), ...)

Arguments

Value

The session, invisibly.

Description

Base-graphics aggregated view of a conflibert_classify result: texts per class with average confidence. For a ggplot2 version use ggplot2::autoplot().

Usage

## S3 method for class 'conflibert_classify'
plot(x, ...)

Arguments

Value

The object, invisibly.

Description

Base-graphics dot chart of metrics per model for a conflibert_compare result. For a ggplot2 version use ggplot2::autoplot().

Usage

## S3 method for class 'conflibert_comparison'
plot(x, ...)

Arguments

Value

The object, invisibly.

Description

Base-graphics test-set confusion matrix for a conflibert_finetune result. For a ggplot2 version use ggplot2::autoplot().

Usage

## S3 method for class 'conflibert_finetune'
plot(x, ...)

Arguments

Value

The object, invisibly.

Description

Base-graphics version of the multilabel plot. With several texts it aggregates (share of texts flagged per category); with a single text it shows that text's category probabilities. For a ggplot2 version use ggplot2::autoplot().

Usage

## S3 method for class 'conflibert_multilabel'
plot(x, ...)

Arguments

Value

The object, invisibly.

Description

Base-graphics aggregated view of a conflibert_ner result: entity counts per type. For a ggplot2 version (including a top-entities view) use ggplot2::autoplot().

Usage

## S3 method for class 'conflibert_ner'
plot(x, ...)

Arguments

Value

The object, invisibly.

Description

Run batched inference with a classifier loaded via conflibert_load. Returns a tibble with predicted class, confidence, and one prob_* column per class.

Usage

## S3 method for class 'conflibert_classifier'
predict(object, text, batch_size = 32, max_seq_len = 512, ...)

Arguments

Value

A tibble with text, class, confidence, and prob_0..prob_{K-1} columns.

Examples

data <- conflibert_example("binary")
dir <- file.path(tempdir(), "my_model")
conflibert_finetune(
  train = data$train, dev = data$dev, test = data$test,
  epochs = 1, save_dir = dir
)
clf <- conflibert_load(dir)
predict(clf, c("Bomb exploded.", "Stock market rose."))

Description

A modern, flat ggplot2 theme used by all of the package's autoplot() methods: no tick marks, no panel box, hairline grid lines only along the data axis, bold left-aligned titles, and generous whitespace. Use it on your own plots to match.

Usage

theme_conflibert(
  base_size = 12,
  base_family = "",
  grid = c("xy", "x", "y", "none")
)

Arguments

Value

A ggplot2 theme object.

Examples

if (requireNamespace("ggplot2", quietly = TRUE)) {
  ggplot2::ggplot(mtcars, ggplot2::aes(wt, mpg)) +
    ggplot2::geom_point(colour = "#0ea5e9", size = 3) +
    theme_conflibert(grid = "y")
}