nano-gpt-data-curation
gpt-5.6-sol
nano-gpt-data-curation — graded by the task's verifier against a measured baseline.
- backend
- modal
- variant
- codex azure_ai/gpt-5.6-sol
- metric
- held-out multi-domain perplexity
- actions
- 261
- duration
- 1h 13m
- recorded
- 2026-08-04
run iddclm-nano-curation__XegPhjw
Score
reward
348.640
Summary
Completed and validated:
- selection.json: 17,616 unique valid IDs; fills exactly 12M tokens.
- curate.py: reproducible quality filtering, deduplication, domain routing, and weighted scheduling.
- claim.md: hypothesis, mechanism, falsification, and transfer.
Final tested dev perplexity: 346.595. The packed submission exactly matches that tested candidate.