Add Velrim absent-field extraction bake-off - #4
Open
plsinvoke wants to merge 1 commit into
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds the Velrim absent-field extraction bake-off: a pre-registered six-arm benchmark of how often document-extraction systems invent values for fields the document does not contain (124 real documents, 2,102 labeled fields from CORD-v2, DeepForm and VRDU; Gemini 2.5 Flash x2, GPT-5.4-mini x2, Mistral Document AI OCR, Velrim).
Disclosure: Velrim ran the comparison and is one of the six systems. The analysis plan was committed publicly before the first paid call, every raw output and request ID is in the repo, and re-scoring needs no API keys.
Entry follows the existing format of this section. Happy to adjust wording or placement.