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
- 273
- duration
- 60m
- recorded
- 2026-08-04
run iddclm-nano-curation__zzFhaz8
Score
reward
338.210
Summary
Completed all deliverables:
- selection.json: 14,831 unique IDs; packs exactly 12M tokens.
- curate.py: deterministic, byte-for-byte reproducible selection.
- claim.md: hypothesis, mechanism, falsification, and transfer.
Final balanced selection achieved dev perplexity 313.591, outperforming tested quality-only and technical-heavy ablations.