Skip to content

[Leaderboard] DVIndex - GPT-6 Astra/high + Jev - 87.68% local Pass@1 - #113

Closed
Legend398 wants to merge 2 commits into
ucbepic:mainfrom
Legend398:dvindex-dab-20260923
Closed

Legend398 wants to merge 2 commits into
ucbepic:mainfrom
Legend398:dvindex-dab-20260923

Conversation

@Legend398

@Legend398 Legend398 commented Sep 23, 2026 •

Copy link
Copy Markdown

DVIndex — Astra/high with Jev classification — 87.68% local Pass@1

Please review this 270-record DVIndex submission and its disclosed complete-question reruns. Local scoring gives 238/270 raw, 0.8768498168 dataset-macro Pass@1, pending your validation.

  • Agent: DVIndex, protected native data tools on OpenCode V2.
  • Backbone: gpt-6-astra, high reasoning; optional helpers use the same model/effort.
  • Classification tool model: Jev 1.13.0 for AGNews Q2/Q3/Q4.
  • Hints: yes, original db_description_withhint.txt.
  • Tuned prompts: yes; generic prompts and dataset-specific guidance are supplied with a complete per-question assignment table.
  • Coverage: all 54 questions, 12 datasets, runs 0–4, 270 selected answer/trace pairs.

Reruns and interruptions

All five trials were rerun and replaced for AGNews Q2/Q3/Q4 and CRM Q2/Q7/Q8/Q12. All 35 rerun outcomes are included; no individual result was selected by correctness. Classification reruns repair a prepared-input launch integration error. CRM reruns use revised dataset method guidance and are disclosed as a changed prompt configuration. The other 235 original pairs remain unchanged.

The original run was paused/resumed across 19–23 September. Real timestamps and failed/cancelled outcomes are retained. One previously performed isolated Yelp Q7 run 2 retry is supplied as supplemental evidence but is not selected; the main result counts its original cancellation as a non-pass. Original superseded recordings are also included for audit.

Files

leaderboard_submissions/dvindex.json is the answer file. The complete evidence archive (download page) contains full traces, original evidence, per-row Jev receipts, prompts, timing, replacement-map.json, configuration fingerprints and local verification. Nothing in this draft asserts that the replacement policy or prompts have already been approved. Please advise if additional evidence or a different treatment is required.

Local results

Dataset Passed Rate
DEPS_DEV_V1 5/10 50.00%
GITHUB_REPOS 20/20 100.00%
PANCANCER_ATLAS 13/15 86.67%
PATENTS 14/15 93.33%
agnews 10/20 50.00%
bookreview 15/15 100.00%
crmarenapro 54/65 83.08%
googlelocal 20/20 100.00%
music_brainz_20k 15/15 100.00%
stockindex 15/15 100.00%
stockmarket 23/25 92.00%
yelp 34/35 97.14%

Scored against unchanged official validators at aae9730; current-main material equality was checked on 23 September. Every selected answer matches its recorded final answer. No manual score overrides or answer rewrites were applied.

Ruiying-Ma added a commit that referenced this pull request Sep 26, 2026
@Ruiying-Ma

Copy link
Copy Markdown
Collaborator

Hi @Legend398! Thanks for the submission and the detailed traces. We have added your results to the leaderboard.

@Ruiying-Ma Ruiying-Ma closed this Sep 26, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants