Repository navigation
Benchmark on real summarization data - #5
Merged
Merged
Conversation
The synthetic Gutenberg benchmark said little about real workloads. The new benchmark scores lead-sentence baselines against references from pinned CNN/DailyMail, XSum and PubMed test files (2,000 pairs each), checks score parity exactly, and reports serial, threaded and Evaluate timings plus a like-for-like comparison with rouge-rust, which is faster on the subset it supports. A manual workflow collects the same numbers on Linux.
Merged
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Replaces the synthetic Gutenberg benchmark with measurements on pinned test files from CNN/DailyMail, XSum and PubMed (2,000 pairs each, lead-sentence baselines as predictions, SHA-256-verified downloads kept outside the repository).
rouge-score0.1.2; the run fails otherwise.rougevsSyntaxSpirits/rougewith default bootstrap aggregation, and peak RSS.rouge-ruston the subset it supports (ROUGE-1/2/L, no stemming): its scores matchrouge-scoreon this sample, and it is faster than this package there.rouge(pltrdy) is timed only as a reference for a different computation.benchmark.ymlruns the same benchmark on Linux on manual dispatch only.On an M1 Pro: 10.6–39.4× faster than
rouge-scoreon one core, 39.7–201× with 8 threads, 2.3–63.5× through Evaluate.