Skip to content
nano-gpt-data-curation

claude-opus-4-8

nano-gpt-data-curation — graded by the task's verifier against a measured baseline.

backend
modal
variant
claude claude-opus-4-8
metric
held-out multi-domain perplexity
actions
37
duration
59m
recorded
2026-08-04

run iddclm-nano-curation__LYgVBSB

Score

reward
379.850
compare against every run on this task →

Summary

All deliverables are complete and verified. Summary:

Result

Held-out perplexity on the dev target: 413.0, vs 459.8 for a random selection — ~10% lower (lower is better).

What I found

The disclosed target is broad (Wikipedia + HQ web + news + technical Q&A) and the budget is tiny (12M tokens, ~6% of the pool). I tested the intuitive approach first — rank documents by resemblance to the target — and every ranked variant was worse than random:

approachPPL
Fluency (target-LM) ranking579
Naive-Bayes target-vs-pool ranking528
random baseline460
junk-filter + dedup + uniform sample (submission)413

Quality-ranking is self-defeating under a broad target + small budget: it concentrates the budget on the single most "fluent" slice and collapses register coverage (vocabulary shrank 63%; the code-bearing technical-Q&A register dropped from 5% → 1%). Coverage, not resemblance, is what a broad target rewards.

The winning recipe (curate.py)

  1. Loose junk filter — drop only clear garbage (non-English by function-word ratio, symbol-soup, single-token spam, numeric dumps, too-short). Thresholds deliberately loose so every genuine register survives.
  2. Near-dup removal — normalized head signature; templated web dupes waste budget.
  3. Uniform (seeded) sampling of survivors to fill the budget — preserves the pool's natural register mix.

Self-contained, runs in ~67s, emits 23,473 valid unique ids (~24M tokens).

Deliverables

  • /workspace/submission/selection.json — 23,473 priority-ordered ids (validated: unique, in-range, >12M tokens)
  • /workspace/submission/curate.py — reproducible, stated-criterion script (reproduces PPL 412.98 exactly)
  • /workspace/claim.md — Hypothesis / Mechanism (predicts diversity & code-register fraction track PPL, confirmed before measuring PPL) / Falsification / Transfer

Transcript