Repository navigation
The typed-graph bench runs twice more on deepseek-flash, and its judge retries a rate limit - #1029
Merged
Merged
Conversation
…e retries a rate limit Two groups of 100 Re-DocRED documents on dev @ 798fb24 with deepseek-flash as the extractor, bge-m3 through OpenRouter and gemini-3.5-flash as the judge: judged precision 86.5% / 89.0% after errata, same-sentence gold recall 12.3% / 13.2% (gemini's fourth run had 17.8% / 17.1%; the extractor finds a third fewer entity pairs), 44k / 48k tokens per document (alignment's thinking under `low`). The errata agent adds without retracting. 0044's status line reads the three thresholds over the fourth and ninth runs and leaves the token threshold as the open decision. The bench's judge call now retries 429 and 5xx with a backoff: four documents in parallel against a per-minute limit lost half of a 200-item sample, and not at random, since a refused call drops a whole document. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: Wayland Yang <wayland0916@gmail.com>
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Two more groups of the typed-graph bench (0044 cut 2 and cut 6) on dev @ 798fb24, 100 Re-DocRED documents each,
--judge 200 --errata, withdeepseek-flashas the extractor (thinking off), bge-m3 through OpenRouter and gemini-3.5-flash as the judge. The bench README gains the ninth run; 0044's status line now reads the three thresholds over the fourth (gemini) and ninth (deepseek) runs.Read against the thresholds: precision passes on both models; same-sentence recall passes on gemini and fails on deepseek-flash, which finds a third fewer entity pairs at extraction; tokens per document fail on both, and the extra on deepseek is its thinking under alignment's fixed
low. The two groups agree to within a point. The open item on 0044 is now the token threshold: restate it per signature, or accept the miss.Bench change: the judge call in
typed.mjsretries 429 and 5xx with a backoff. Four documents in parallel against OpenRouter's per-minute limit lost half of each 200-item sample, and not at random (a refused call drops a whole document); both bases were re-scored with the retry and judged in full.Records:
~/code/utopia-bench-runs/0044/on the bench machine (not committed).🤖 Generated with Claude Code