Skip to contents

Changelog

Source: NEWS.md
  • fuzzylink() now supports Anthropic Claude (e.g. 'claude-sonnet-4-5-20250929') and Mistral (e.g. 'mistral-large-latest') models in addition to OpenAI.
  • Updated default model to gpt-5.2.
  • Active learning loop now reports rolling gradient and stopping threshold in real time.
  • Fixed a potential edge case in get_cutoff() where NaN F1 scores could cause silent failures on datasets with no true matches.
  • Removed unused internal functions (get_training_set(), hand_label(), estimate_tokens()).
  • Fixed a bug in check_match() where verbose LLM responses (e.g. from Claude or Mistral) could cause match labels to be parsed incorrectly. Labels are now normalized by extracting the first word of each response.
  • Patched check_match() to return labels from legacy OpenAI models.

CRAN release: 2026-01-23

  • API calls for language model prompts are now passed through the ellmer package, improving speed, error handling, and rate limits.

CRAN release: 2025-08-29

  • Updated documentation with publication DOI

CRAN release: 2025-08-18

  • Patched a bug introduced in 0.2.3, crashing the algorithm when there are no exact matches.
  • The algorithm now omits exact matches from the training set during the active learning loop. This avoids the superfluous step of labeling exact matches with the language model, and patches a bug wherein too many exact matches caused the loop to terminate prematurely.
  • The algorithm now drops missing observations from dfA and dfB with a warning.

CRAN release: 2025-06-14

  • Updated package documentation and console messages.
  • Initial CRAN submission.