Repository navigation
Publish all 85 queries - #46
Merged
Merged
Conversation
`tools/make_cuad_labels.py <root> --extend-published` writes the files that add the 18 CUAD label sets to the published ground truth: label sets whose ids come from the published corpus, one new collection per active collection of that corpus with the CUAD label sets added, and the pointers naming the new collections. The corpus files are not written, since the benchmark builds the CUAD tables from their sources. `--upload` copies the files to the public bucket with the machine's AWS credentials. The local mirror the tool wrote before stays as the first form; it reuses the published corpus id with a two-table manifest, so it must not be uploaded. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0121QACZnjC3Gv6AnWnWSBB8
The tool now has one job: write the files that add the CUAD label sets to the published ground truth, for every published scale factor, and upload them with `--upload`. The local mirror mode is gone, since the benchmark reads the labels from the bucket. The upload uses boto3, which resolves the bucket's region and credentials the way the AWS CLI does; pyarrow's S3 filesystem failed the bucket check with an HTTP 400 on a machine with working credentials. After uploading, the tool reads the pointers back from the bucket to confirm them. The CUAD queries lose `labels_pending`, so `queries()` returns 65 queries over six document collections. The tests and docs that stated 50 now state 65. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0121QACZnjC3Gv6AnWnWSBB8
shreyashankar
changed the title
The trace and sales queries were marked labels_pending although their labels were in the bucket, on newer corpora than PUBLISHED_CORPORA named. The CUAD tables were in no corpus, so quail_b.run could not load them. The labels tool now writes, for each scale factor, a corpus with the trace corpus tables plus the CUAD tables, the CUAD label sets, and a collection that reuses the existing label sets. The upload copies the unchanged tables inside the bucket. PUBLISHED_CORPORA names the new corpora, and labels_pending is removed: every query has labels at every scale factor. The new filters and joins get selectivity estimates from the sf=0.1 labels. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LnBPDnvaDsBDWZofvX25DU
shreyashankar
changed the title
The download cache copied each S3 file through a random access file, which sends one HTTP range request per 64 KB read. Downloads ran at about 0.2 MB/s, so the scale factor 1.0 tables alone took about 20 minutes. The cache now reads a sequential stream, reuses one S3 filesystem, and load_benchmark fetches tables in parallel. A cold load at scale factor 1.0 takes 80 s and a warm one 6 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LnBPDnvaDsBDWZofvX25DU
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
2 participants
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.
All 85 queries now have published labels at every scale factor. Nothing is pending.
PUBLISHED_CORPORAnames three new corpora. Each one has the tables of the trace corpus for its scale factor plus the CUAD tables (contracts,clause_types). Before this change, the CUAD tables were in no corpus, soquail_b.runcould not load them.labels_pendingis removed. The trace and sales queries were marked pending even though their labels were already in the bucket.tools/make_cuad_labels.pywrites the new corpora and collections.--uploadcopies the unchanged tables inside the bucket.load_benchmarkfetches tables in parallel. At scale factor 1.0, a cold load takes 80 s and a warm one 6 s.c_e380e53469a6a62de83a452fbc51a73egt_349c059f13738eaeb76d230361f3c219c_38414b12549ab6243397fd41522e67adgt_1d3d18a553eb491a99b425524c21ea16c_3a8712778b5fe0938abe729770c4d23dgt_b81d2437aaa5080fd54c81d82833a3e5These files are already uploaded. Against the bucket, at each scale factor:
Ruff, the long-string check, vulture, and pytest pass.
🤖 Generated with Claude Code
https://claude.ai/code/session_01LnBPDnvaDsBDWZofvX25DU