code/derived-data.R

library(tidyverse)
library(scales)
library(ggrepel)
library(RColorBrewer)
library(optparse)

option_list <- list(
  make_option(c("-c", "--core"),
              type = "character", default = "core",
              help = "Directory containing the Google Sheets downloads.", metavar = "character"
              ),
  make_option(c("-o", "--output"),
              type = "character", default = ".",
              help = "Output directory.", metavar = "character"
              )
)

                                        # Parse command line arguments
opt_parser <- OptionParser(option_list = option_list)
opt <- parse_args(opt_parser)

                                        # Check if required arguments are provided
if (is.null(opt$core)) {
  stop("Error: Sheets directory (-s, --sheets) is required")
}
if (is.null(opt$output)) {
  stop("Error: Output directory (-o, --output) is required")
}

core_dir <- opt$core
output_dir <- opt$output

                                        # Load the CSV source files.
################
## Core datasets
################
model_hierarchy <- c(
  "model_family",
  "model_version",
  "model_variant",
  "model_subvariant",
  "model_base"
)

core_models <- read_csv(str_glue("{core_dir}/models.csv")) |>
  mutate(
    model_base = if_else(model_is_base == TRUE, "base", NA_character_),
    model_old_id = model_id,
    ) |>
  unite(
    col = "model_id",
    all_of(model_hierarchy),
    sep = "_",
    remove = FALSE,
    na.rm = TRUE # this skips fields that are NA.
  ) |>
  mutate(
    model_id = str_to_lower(model_id),
    model_id = str_replace_all(model_id, fixed("."), "_")
  )

core_highlights <- core_models |>
                                        # Filter here to exclude base models.
                                        # filter(model_is_base == FALSE) |>
  inner_join(read_csv(str_glue("{core_dir}/highlights.csv")), by = join_by(model_old_id == model_id)) |>
  select(
    model_id,
    benchmark_id,
    prescribed_competency,
    prescribed_category,
    model_is_base,
    model_published_at,
    model_access,
    organization_name,
    organization_sector,
    organization_domain,
    all_of(model_hierarchy),
    ) |>
  group_by(
    benchmark_id, model_id,
    ) |>
  summarise(
                                        # Get unique, non-NA competencies, and combine them with ' && '
    prescribed_competency = paste(unique(na.omit(prescribed_competency)), collapse = " && "),
                                        # Get unique, non-NA categories, and combine them with ' && '
    prescribed_category = paste(unique(na.omit(prescribed_category)), collapse = " && "),
    model_published_at = first(model_published_at),
    model_access = first(na.omit(model_access)),
    organization_name = first(na.omit(organization_name)),
    organization_sector = first(na.omit(organization_sector)),
    organization_domain = first(na.omit(organization_domain)),
    .groups = "keep"
  ) |>
  ungroup()

core_models <- core_highlights |>
  group_by(model_id) |>
  mutate(model_published_at = min(model_published_at)) |>
  ungroup() |>
  select(
    model_id,
    model_published_at,
    model_access,
    organization_name,
    organization_sector,
    organization_domain,
    ) |>
  distinct()

core_benchmarks <- core_highlights |>
  left_join(read_csv(str_glue("{core_dir}/benchmarks.csv")), by = "benchmark_id") |>
  select(
    benchmark_id,
    benchmark_name,
    benchmark_full_name,
    paper_id,
    paper_href,
    paper_published_at,
    ) |>
  distinct()

core_highlights <- core_highlights |>
  select(
    model_id,
    benchmark_id,
    prescribed_competency,
    prescribed_category,
    ) |>
  distinct()

core_categories <- read_csv(str_glue("{core_dir}/categories.csv"))
core_categorizations <- read_csv(str_glue("{core_dir}/categorizations.csv")) |>
  left_join(core_categories, by = "benchmark_category")
core_affiliations <- read_csv(str_glue("{core_dir}/affiliations.csv"))
core_knowledge_subjects <- read_csv(str_glue("{core_dir}/knowledge-subjects.csv"))

###########
## Rankings
###########
rankings_overall <- core_highlights |>
  left_join(core_models, by = "model_id") |>
  select(benchmark_id, model_id, organization_name) |>
                                        # Calculate the totals of all models and publishers.
  mutate(
    total_publishers = n_distinct(organization_name),
    total_models = n_distinct(model_id)
  ) |>
  summarise(
    n_publishers = n_distinct(organization_name),
    n_models = n_distinct(model_id),
                                        # Carry the totals forward (they are the same for every row, so first() works)
    total_publishers = first(total_publishers),
    total_models = first(total_models),
    .by = benchmark_id
  ) |>
  mutate(
    score = sqrt(n_publishers * n_models),
    rank = dense_rank(desc(score)),
    p_publishers = n_publishers / total_publishers,
    p_models = n_models / total_models,
    ) |>
  select(-total_publishers, -total_models) |>
  select(benchmark_id, rank, ends_with("publishers"), ends_with("models"), score)

rankings_access <- core_highlights |>
  left_join(core_models, by = "model_id") |>
  select(benchmark_id, model_id, model_access, organization_name) |>
                                        # Calculate the totals of all models and publishers grouped by model_access.
  mutate(
    total_models = n_distinct(model_id),
    .by = model_access
  ) |>
  summarise(
    n_publishers = n_distinct(organization_name),
    n_models = n_distinct(model_id),
                                        # Carry the totals forward (they are the same for every row, so first() works)
    total_models = first(total_models),
    .by = c(benchmark_id, model_access)
  ) |>
  group_by(model_access) |>
  mutate(
    score = sqrt(n_publishers * n_models),
    rank = dense_rank(desc(score)),
    p_models = n_models / total_models,
    ) |>
  ungroup() |>
  select(-total_models) |>
  pivot_wider(
    id_cols = benchmark_id,
    names_from = model_access,
    values_from = c(rank, n_models, p_models, score),
    names_glue = "{.value}_{str_to_lower(str_replace_all(model_access, ' ', '_'))}"
  ) |>
  select(benchmark_id, ends_with("closed"), ends_with("open-weight"), ends_with("open-source"))

rankings_domain <- core_highlights |>
  left_join(core_models, by = "model_id") |>
  select(benchmark_id, model_id, organization_domain, organization_name) |>
                                        # Calculate the totals of all models and publishers grouped by organization_domain.
  mutate(
    total_publishers = n_distinct(organization_name),
    .by = organization_domain
  ) |>
  summarise(
    n_publishers = n_distinct(organization_name),
    n_models = n_distinct(model_id),
                                        # Carry the totals forward (they are the same for every row, so first() works)
    total_publishers = first(total_publishers),
    .by = c(benchmark_id, organization_domain)
  ) |>
  group_by(organization_domain) |>
  mutate(
    score = sqrt(n_publishers * n_models),
    rank = dense_rank(desc(score)),
    p_publishers = n_publishers / total_publishers,
    ) |>
  ungroup() |>
  select(-total_publishers) |>
  pivot_wider(
    id_cols = benchmark_id,
    names_from = organization_domain,
    values_from = c(rank, n_publishers, p_publishers, score),
    names_glue = "{.value}_{str_to_lower(str_replace_all(organization_domain, ' ', '_'))}"
  ) |>
  select(benchmark_id, ends_with("west"), ends_with("china"))

rankings <- rankings_overall |>
  left_join(rankings_domain, by = "benchmark_id") |>
  left_join(rankings_access, by = "benchmark_id") |>
  select(
    benchmark_id,
    rank,
    ends_with("models"),
    ends_with("publishers"),
    ends_with("publishers_west"),
    ends_with("publishers_china"),
    ends_with("models_closed"),
    ends_with("models_open-weight"),
    ends_with("models_open-source"),
    starts_with("rank_"),
    contains("score"),
    ) |>
  arrange(rank) |>
  write_csv(str_glue("{output_dir}/derived/rankings.csv"), na = "")

## Top 15 Most Popular Benchmarks: Table 4
rankings_top_15 <- rankings |>
  slice_min(order_by = rank, n = 15, with_ties = TRUE) |>
  write_csv(str_glue("{output_dir}/derived/rankings-top-15.csv"), na = "")

##############
                                        # Affiliations
##############
benchmark_creators <- core_affiliations |>
  inner_join(core_benchmarks, by = "paper_id", relationship = "many-to-many")

## Affiliation of Benchmark Creators: Table 1
affiliations_overall <- benchmark_creators |>
  count(organization_sector) |>
  mutate(p = n / sum(n))

affiliations_west <- benchmark_creators |>
  filter(organization_domain == "west") |>
  count(organization_sector) |>
  mutate(p = n / sum(n))

affiliations_china <- benchmark_creators |>
  filter(organization_domain == "china") |>
  count(organization_sector) |>
  mutate(p = n / sum(n))

affiliations_overall_2025 <- benchmark_creators |>
  filter(year(paper_published_at) == "2025") |>
  count(organization_sector) |>
  mutate(p = n / sum(n))

affiliations_west_2025 <- benchmark_creators |>
  filter(year(paper_published_at) == "2025" & organization_domain == "west") |>
  count(organization_sector) |>
  mutate(p = n / sum(n))

affiliations_china_2025 <- benchmark_creators |>
  filter(year(paper_published_at) == "2025" & organization_domain == "china") |>
  count(organization_sector) |>
  mutate(p = n / sum(n))

                                        # We join the separate affiliation tibbles into a single data structure
                                        # containing all columns joined on the organization_sector column.
list(
  overall      = affiliations_overall,
  overall_2025 = affiliations_overall_2025,
  west         = affiliations_west,
  west_2025    = affiliations_west_2025,
  china        = affiliations_china,
  china_2025   = affiliations_china_2025
) |>
                                        # imap iterates over the list (.x is the tibble, .y is the list name)
  imap(~ rename_with(.x, function(col) paste0(col, "_", .y), c(n, p))) |>
                                        # reduce sequentially joins all tibbles in the list
  reduce(~ full_join(.x, .y, by = "organization_sector")) |>
  write_csv(str_glue("{output_dir}/derived/affiliations-benchmark-creators.csv"), na = "")

## Multiple affiliations of benchmark authors: Table 2
n_benchmark_creators <- benchmark_creators |>
  distinct(author_name, paper_id) |>
  nrow()
n_benchmark_creators_multiple <-
  benchmark_creators |>
  add_count(author_name, paper_id) |>
  filter(n >= 2) |>
  distinct(author_name, paper_id) |>
  nrow()

tibble(
  n_benchmark_creators = n_benchmark_creators,
  n_benchmark_creators_multiple = n_benchmark_creators_multiple,
  p_benchmark_creators_multiple = n_benchmark_creators_multiple / n_benchmark_creators
) |>
  write_csv(str_glue("{output_dir}/derived/benchmark-creators.csv"), na = "")

benchmark_creators |>
                                        # Only proceed with authors that appear more than once.
  add_count(author_name, paper_id) |>
  filter(n >= 2) |>
                                        # Create combined sector labels. Make sure the sectors are always appearing in
                                        # the same order.
  group_by(author_name, paper_id) |>
  arrange(organization_sector, .by_group = TRUE) |>
  summarize(
    combination = paste(unique(organization_sector), collapse = " & "),
    .groups = "drop",
    ) |>
                                        # Count the combination labels and produce final output.
  count(combination) |>
  mutate(p = n / sum(n)) |>
  arrange(desc(p)) |>
  write_csv(str_glue("{output_dir}/derived/affiliations-benchmark-creators-multiple.csv"), na = "")

                                        # Affiliations per benchmark
benchmark_affiliations <- benchmark_creators |>
  group_by(benchmark_id) |>
  count(organization_sector) |>
  mutate(p = n / sum(n)) |>
  ungroup() |>
  select(-n) |>
  write_csv(str_glue("{output_dir}/derived/affiliations-benchmark.csv"), na = "")

benchmark_affiliations |>
  pivot_wider(
    names_from = organization_sector,
    values_from = p,
    names_glue = "{.value}_{organization_sector}",
    ) |>
  relocate(`p_NA`, .after = last_col()) |>
  write_csv(str_glue("{output_dir}/derived/affiliations-benchmark-wide.csv"), na = "")

##############
                                        # Competencies
##############
## Evaluated Competencies in the Top 15 Benchmarks: Table 3
top_15_categories <- rankings_top_15 |>
  left_join(core_categorizations, by = "benchmark_id")

evaluated_competencies <- top_15_categories |>
  count(benchmark_category) |>
  mutate(p = n / sum(n))

evaluated_meta_competencies <- top_15_categories |>
  count(benchmark_meta_category) |>
  mutate(p = n / sum(n))

top_15_categories |>
  left_join(core_benchmarks, by = "benchmark_id") |>
  group_by(benchmark_category) |>
  arrange(benchmark_name, .by_group = TRUE) |>
  summarize(
    benchmarks = paste(unique(benchmark_name), collapse = ", "),
    .groups = "drop",
    ) |>
  left_join(evaluated_competencies, by = "benchmark_category") |>
  arrange(desc(p)) |>
  write_csv(str_glue("{output_dir}/derived/evaluated-competencies.csv"), na = "")

top_15_categories |>
  left_join(core_benchmarks, by = "benchmark_id") |>
  group_by(benchmark_meta_category) |>
  arrange(benchmark_name, .by_group = TRUE) |>
  summarize(
    benchmarks = paste(unique(benchmark_name), collapse = ", "),
    .groups = "drop",
    ) |>
  left_join(evaluated_meta_competencies, by = "benchmark_meta_category") |>
  arrange(desc(p)) |>
  write_csv(str_glue("{output_dir}/derived/evaluated-meta-categories.csv"), na = "")

## Yearly breakdown of benchmark competencies: Table 5
core_categorizations |>
  left_join(core_benchmarks, by = "benchmark_id") |>
  mutate(year = year(paper_published_at)) |>
  count(benchmark_meta_category, benchmark_category, year) |>
  group_by(benchmark_category) |>
  mutate(p = n / sum(n)) |>
  pivot_wider(
    names_from = year,
    values_from = c(n, p),
    names_glue = "{.value}_{year}",
    names_sort = TRUE
  ) |>
  arrange(benchmark_meta_category, benchmark_category) |>
  write_csv(str_glue("{output_dir}/derived/evaluated-competencies-by-year.csv"), na = "")

## Prescribed competencies by model publishers within "Coding" benchmarks: Figure 1
prescribed_competencies_coding <- core_highlights |>
  left_join(rankings, by = "benchmark_id") |>
  inner_join(core_categorizations, by = "benchmark_id", relationship = "many-to-many") |>
  left_join(core_benchmarks, by = "benchmark_id") |>
  filter(benchmark_category == "Coding") |>
  select(benchmark_id, benchmark_name, model_id, prescribed_category, rank) |>
  write_csv(str_glue("{output_dir}/derived/prescribed-competencies-coding.csv"), na = "")

## Prescribed competencies by model publishes for the LiveCodeBench benchmark: RQ2
prescribed_competencies_coding |>
  filter(benchmark_id == "lcb") |>
  count(prescribed_category) |>
  distinct(prescribed_category, n) |>
  mutate(p = n / sum(n)) |>
  write_csv(str_glue("{output_dir}/derived/prescribed-competencies-lcb.csv"), na = "")

top_15_coding_benchmarks <- prescribed_competencies_coding |>
  distinct(benchmark_name, rank) |>
  slice_min(order_by = rank, n = 5, with_ties = FALSE) |>
  pull(benchmark_name)

prescribed_competencies_coding |>
  filter(!is.na(prescribed_category)) |>
  filter(benchmark_name %in% top_15_coding_benchmarks) |>
  mutate(prescribed_category = factor(prescribed_category)) |>
  mutate(benchmark_name = fct_reorder(benchmark_name, rank, .desc = TRUE)) |>
  count(benchmark_name, prescribed_category) |>
  complete(benchmark_name, prescribed_category) |>
  ggplot(aes(x = prescribed_category, y = benchmark_name)) +
  geom_tile(
    data = \(x) filter(x, is.na(n)),
    fill = "grey95", color = "white", linewidth = 0.5
  ) +
  geom_tile(
    data = \(x) filter(x, !is.na(n)),
    aes(fill = n), color = "white", linewidth = 0.5
  ) +
  geom_text(
    data = \(x) filter(x, !is.na(n)),
    aes(label = n, color = n > max(n, na.rm = TRUE) / 2),
    size = 4, show.legend = FALSE
  ) +
  scale_color_manual(values = c("black", "white")) +
  scale_fill_viridis_c(option = "plasma", name = "Number of\nHighlights", direction = -1) +
  labs(
    x = "Prescribed Category",
    y = NULL,
    title = "Prescribed Competencies by Model Publisher",
    subtitle = "Within Top 5 Coding Benchmarks",
    caption = 'From: "Unsteady Metrics and Benchmarking Cultures of AI Model Builders", submitted to FAccT 2026'
  ) +
  theme_minimal() +
  theme(
    axis.text.x = element_text(angle = 45, hjust = 1),
    panel.grid = element_blank()
  ) +
  coord_fixed(ratio = 1)

ggsave(
  filename = str_glue("{output_dir}/figures/prescribed-competencies-coding.png"),
  width = 10, height = 6, dpi = 300,
  bg = "white"
)

## Prescribed competencies by model publishers within top 15 "Reasoning and knowledge" benchmarks: Figure 4
prescribed_competencies_reasoning_knowledge <- rankings |>
  slice_min(order_by = rank, n = 15, with_ties = TRUE) |>
  inner_join(core_highlights, by = "benchmark_id") |>
  filter(rank <= 15) |>
  inner_join(core_categorizations, by = "benchmark_id", relationship = "many-to-many") |>
  left_join(core_benchmarks, by = "benchmark_id") |>
  filter(benchmark_category == "Reasoning and knowledge") |>
  select(benchmark_id, benchmark_name, model_id, prescribed_category, rank) |>
  write_csv(str_glue("{output_dir}/derived/prescribed-competencies-reasoning-knowledge.csv"), na = "")

prescribed_competencies_reasoning_knowledge |>
  filter(!is.na(prescribed_category)) |>
                                        # For consistency, we remove MMMLU from the graph, since it is also missing in
                                        # other tables/figures mentioning the Top 5 "Reasoning and knowledge"
                                        # benchmarks. See footnote in paper.
  filter(benchmark_id != "mmmlu") |>
  mutate(prescribed_category = factor(prescribed_category)) |>
  mutate(benchmark_name = fct_reorder(benchmark_name, rank, .desc = TRUE)) |>
  count(benchmark_name, prescribed_category) |>
  complete(benchmark_name, prescribed_category) |>
  ggplot(aes(x = prescribed_category, y = benchmark_name, fill = n)) +
  geom_tile(
    data = \(x) filter(x, is.na(n)),
    fill = "grey95", color = "white", linewidth = 0.5
  ) +
  geom_tile(
    data = \(x) filter(x, !is.na(n)),
    aes(fill = n), color = "white", linewidth = 0.5
  ) +
  geom_text(
    data = \(x) filter(x, !is.na(n)),
    aes(label = n, color = n > max(n, na.rm = TRUE) / 2),
    size = 4, show.legend = FALSE
  ) +
  scale_color_manual(values = c("black", "white")) +
  scale_fill_viridis_c(option = "plasma", name = "Number of\nHighlights", direction = -1) +
  labs(
    x = "Prescribed Category",
    y = NULL,
    title = "Prescribed Competencies by Model Builders",
    subtitle = "Within Reasoning and Knowledge Category of the Top 15 Benchmarks",
    caption = 'From: "Unsteady Metrics and Benchmarking Cultures of AI Model Builders", submitted to FAccT 2026',
    ) +
  theme_minimal() +
  theme(
    axis.text.x = element_text(angle = 45, hjust = 1),
    panel.grid = element_blank() # Removes default gridlines behind tiles
  ) +
  coord_fixed(ratio = 1)

ggsave(
  filename = str_glue("{output_dir}/figures/prescribed-competencies-reasoning-knowledge.png"),
  width = 10, height = 6, dpi = 300,
  bg = "white"
)

                                        # Percentage of Unique Models for top 15 Reasoning and knowledge
reasoning_and_knowledge_benchmark_ids <- prescribed_competencies_reasoning_knowledge |>
  distinct(benchmark_id) |>
  pull(benchmark_id)

core_highlights |>
  summarize(
    n = n_distinct(model_id[benchmark_id %in% reasoning_and_knowledge_benchmark_ids]),
    p = (n / n_distinct(model_id))
  ) |>
  write_csv(str_glue("{output_dir}/derived/reasoning-knowledge-unique-models.csv"), na = "")

#####################
## Benchmark Adoption
#####################
## Yearly releases of becnhmarks by year: RQ4
core_benchmarks |>
  mutate(year = year(paper_published_at)) |>
  count(year) |>
  mutate(p = n / sum(n)) |>
  distinct(year, n, p) |>
  arrange(year) |>
  write_csv(str_glue("{output_dir}/derived/benchmark-releases-by-year.csv"), na = "")

## Adoption of benchmarks released in 2025: Figure 2
highlights_2025 <- core_highlights |>
  left_join(core_benchmarks, by = "benchmark_id") |>
  filter(year(paper_published_at) == "2025") |>
  left_join(core_models, by = "model_id") |>
                                        # Calculate the 'T-Zero' relative time
                                        # This normalizes all benchmarks to start at Month 0
  mutate(months_since_release = interval(paper_published_at, model_published_at) %/% months(1)) |>
                                        # Filter out observations that happened before official pub
  filter(months_since_release >= 0) |>
  group_by(benchmark_id, benchmark_name, months_since_release) |>
  summarise(monthly_count = n(), .groups = "drop_last") |>
                                        # Calculate cumulative sum for the S-Curve
  mutate(cumulative_popularity = cumsum(monthly_count)) |>
                                        # Calculate Velocity (Growth rate)
                                        # This shows how many new 'mentions' happen per month
  mutate(velocity = cumulative_popularity - lag(cumulative_popularity, default = 0)) |>
  write_csv(str_glue("{output_dir}/derived/highlights-2025.csv"), na = "")

top_highlights_2025 <- highlights_2025 |>
  group_by(benchmark_name) |>
  summarize(max_pop = max(cumulative_popularity)) |>
  slice_max(max_pop, n = 5) |>
  pull(benchmark_name)

other_highlights_2025 <- c("ARC-AGI-2", "Tau2-bench", "Creative Writing")

highlight_palette <- setNames(
  c(viridis::plasma(5), "#D3D3D3"), # 5 colors + light grey
  c(top_highlights_2025, "Other")
)

highlights_2025_label_data <- highlights_2025 |>
  group_by(benchmark_name) |>
  filter(months_since_release == max(months_since_release))

highlights_2025_label_data_filtered <- highlights_2025_label_data |>
  filter(benchmark_name %in% top_highlights_2025)

highlights_2025_other_label_data_filtered <- highlights_2025_label_data |>
  filter(benchmark_name %in% other_highlights_2025)

highlights_2025_plot_data <- highlights_2025 |>
  mutate(highlight_color = case_when(
           benchmark_name %in% top_highlights_2025 ~ benchmark_name,
           benchmark_name %in% other_highlights_2025 ~ "Secondary",
           TRUE ~ "Other"
         ))

ggplot(
  highlights_2025_plot_data,
  aes(
    x = months_since_release,
    y = cumulative_popularity,
    group = benchmark_name, color = highlight_color,
    )
) +
                                        # Draw the grey lines first, then the colored lines on top
  geom_line(data = filter(highlights_2025_plot_data, highlight_color == "Other"), size = 0.7, alpha = 0.6) +
  geom_line(data = filter(highlights_2025_plot_data, highlight_color == "Secondary"), size = 1.2, alpha = 0.6) +
  geom_line(data = filter(highlights_2025_plot_data, highlight_color != "Other"), size = 1.2, alpha = 1) +

                                        # Apply our custom palette
  scale_color_manual(values = highlight_palette) +
  geom_text_repel(
    data = highlights_2025_label_data_filtered,
    aes(label = benchmark_name, x = months_since_release, y = cumulative_popularity),
    inherit.aes = FALSE,
    hjust = 0,
    nudge_x = 0.5,
    direction = "y",
    segment.color = "grey"
  ) +
  geom_text_repel(
    data = highlights_2025_other_label_data_filtered,
    aes(label = benchmark_name, x = months_since_release, y = cumulative_popularity),
    inherit.aes = FALSE,
    hjust = 0,
    nudge_x = 0.5,
    direction = "y",
    segment.color = "grey"
  ) +
  scale_x_continuous(expand = expansion(mult = c(0.05, 0.3))) +
  theme_minimal() +
  theme(legend.position = "none") +
  labs(
    x = "Months Since Release",
    y = "Cumulative Highlights",
    title = "Adoption of Benchmarks Released in 2025",
    caption = 'From: "Unsteady Metrics and Benchmarking Cultures of AI Model Builders", submitted to FAccT 2026',
    )

ggsave(
  filename = str_glue("{output_dir}/figures/highlights-2025.png"),
  width = 10, height = 6, dpi = 300,
  bg = "white"
)

## Highlights of selected competencies by model builders: Figure 3
category_names <- c(
  "Coding",
  "Math",
  "Reasoning and knowledge",
  "Strategic planning",
  "Tool orchestration"
)

highlighted_competencies <- core_categorizations |>
  left_join(core_highlights, by = "benchmark_id", relationship = "many-to-many") |>
  left_join(core_benchmarks, by = "benchmark_id") |>
  left_join(core_models, by = "model_id") |>
  mutate(month = floor_date(model_published_at, "month")) |>
  count(month, benchmark_meta_category, benchmark_category) |>
  write_csv(str_glue("{output_dir}/derived/highlighted-competencies-by-month.csv"), na = "")

highlighted_competencies |>
  filter(benchmark_category %in% category_names) |>
  ggplot(aes(x = month, y = n, color = benchmark_category)) +
  geom_line(alpha = 0.5) +
  geom_point(alpha = 0.5) +
  geom_smooth(method = "lm", se = FALSE, size = 1.2) +
  scale_color_viridis_d(option = "plasma") +
  scale_fill_viridis_d(option = "plasma") +
  scale_x_date(date_labels = "%b %Y", date_breaks = "1 month") +
  labs(
    x = "Month of Model Publication",
    y = "Number of Highlights",
    color = "Assigned Competency",
    title = "Highlights of Selected Competencies by Model Builders",
    caption = 'From: "Unsteady Metrics and Benchmarking Cultures of AI Model Builders", submitted to FAccT 2026'
  ) +
  guides(color = guide_legend(ncol = 1)) +
  theme_minimal() +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))

ggsave(
  filename = str_glue("{output_dir}/figures/highlighted-competencies-by-month.png"),
  width = 10, height = 6, dpi = 300,
  bg = "white"
)

                                        # Highlighted competencies: Figure 5 (Appendix)
palette_families <- c(
  "Blues", "Reds", "Greens",
  "Purples", "Oranges", "Greys",
  "YlGn", "BuPu"
)

highlighted_competencies_custom_colors <- highlighted_competencies |>
  distinct(benchmark_meta_category, benchmark_category) |>
  arrange(benchmark_meta_category, benchmark_category) |>
  group_by(benchmark_meta_category) |>
  mutate(
    pal_name = palette_families[cur_group_id()],
    color = colorRampPalette(brewer.pal(max(3, n()), pal_name[1]))(n())
  ) |>
  ungroup() |>
  select(benchmark_category, color) |>
  deframe()

highlighted_competencies |>
                                        # Crucial: Order the factors so the legend naturally groups the meta-categories together
  arrange(benchmark_meta_category, benchmark_category) |>
  mutate(benchmark_category = fct_inorder(benchmark_category)) |>
  ggplot(aes(x = month, y = n, color = benchmark_category, fill = benchmark_category)) +
  geom_line(alpha = 0.5) +
  geom_point(alpha = 0.5) +
  geom_smooth(method = "lm", se = FALSE, linewidth = 1.2) +
  facet_wrap(~benchmark_meta_category, scales = "free_y") +
  scale_color_manual(values = highlighted_competencies_custom_colors) +
  scale_fill_manual(values = highlighted_competencies_custom_colors) +
  scale_x_date(date_labels = "%b %Y", date_breaks = "3 month") +
  labs(
    x = "Month of Model Publication",
    y = "Number of Highlights",
    color = "Assigned Competency",
    fill = "Assigned Competency",
    title = "Highlights of Competencies by Model Builders",
    caption = 'From: "Unsteady Metrics and Benchmarking Cultures of AI Model Builders", submitted to FAccT 2026'
  ) +
  guides(color = guide_legend(
           ncol = 5,
           byrow = TRUE,
           title.position = "top",
           override.aes = list(linewidth = 2)
         )) +
  theme_minimal() +
  theme(
    axis.text.x = element_text(angle = 45, hjust = 1),
    legend.position = "bottom",
    legend.text = element_text(size = 8),
    legend.key.width = unit(1, "cm"),
    legend.box.margin = margin(t = 10)
  )

ggsave(
  filename = str_glue("{output_dir}/figures/highlighted-competencies-by-month-facets.png"),
  width = 10, height = 10, dpi = 300,
  bg = "white"
)

## Distribution of benchmark adoption: Table 6
rankings |>
  select(starts_with("n_")) |>
  pivot_longer(
    cols = everything(),
    names_to = "column_name",
    values_to = "highlights"
  ) |>
  filter(!is.na(highlights)) |>
  mutate(column_name = str_remove(column_name, "^n_")) |>
  count(column_name, highlights) |>
  group_by(column_name) |>
  mutate(p = (n / sum(n))) |>
  ungroup() |>
  pivot_wider(
    names_from = column_name,
    values_from = c(n, p),
    names_vary = "slowest"
  ) |>
  select(
    highlights,
    ends_with("_publishers"),
    ends_with("_publishers_west"),
    ends_with("_publishers_china"),
    ends_with("_models"),
    ends_with("_models_closed"),
    ends_with("_models_open-weight"),
    ends_with("_models_open-source"),
    ) |>
  arrange(highlights) |>
  write_csv(str_glue("{output_dir}/derived/benchmark-adoption.csv"), na = "")

###################################
## Knowledge and reasoning subjects
###################################
## Distribution of subjects covered in "Reasoning and Knowledge" benchmarks: Table 8
core_knowledge_subjects |>
  count(field, wt = n, sort = TRUE) |>
  mutate(p = n / sum(n)) |>
  write_csv(str_glue("{output_dir}/derived/subjects-covered.csv"), na = "")

## Breakdown of Science Questions by discipline: Table 9
core_knowledge_subjects |>
  filter(!is.na(science_discipline)) |>
  count(science_discipline, wt = n, sort = TRUE) |>
  mutate(p = n / sum(n)) |>
  write_csv(str_glue("{output_dir}/derived/science-subjects-covered.csv"), na = "")