library(tidyverse)

main_csv <-"~/analysis_data/100325_unified_phab.csv"
main_df <- read.csv(main_csv, header = TRUE) 

set.seed(1893) # For reproducibility

sampled_df <- main_df %>%
  group_by(source) %>%
  mutate(sampled_TaskPHID = TaskPHID %in% sample(unique(TaskPHID), 40)) %>%
  ungroup() %>%
  filter(sampled_TaskPHID) %>%
  select(-sampled_TaskPHID) 

labeling_sampled_df <- sampled_df %>%
  group_by(source) %>%
  mutate(
    verification_sample = if_else(
      TaskPHID %in% sample(unique(TaskPHID), min(8, length(unique(TaskPHID)))), 1L, 0L
    )
  ) %>%
  ungroup()
  
sentence_level_sample <- labeling_sampled_df |>
  mutate(
    cleaned_sentences = lapply(
      olmo_cleaned_sentences,
      function(x) {
        m <- str_match_all(x, "'([^']*)'|\"([^\"]*)\"")[[1]]
        # Both columns 2 and 3 may exist, but only one will be filled for each match
        vals <- c(m[,2], m[,3])
        vals <- vals[vals != "" & !is.na(vals)]
        return(vals)
      }
    )
  )|>
  unnest(cleaned_sentences)|>
  filter(cleaned_sentences != ", ") |>
  select(-olmo_sentence_categories, -starts_with("normalized"), -contains("gerrit"))


table(sentence_level_sample$verification_sample)
(nrow(sentence_level_sample) / 293) * 1.5
length(unique(sentence_level_sample$TaskPHID))

write.csv(sentence_level_sample, "100625_human_info_sample.csv", row.names = FALSE)