Datasets:

caiotheodoro
/

plumb

license: apache-2.0
task_categories:
  - text-generation
language:
  - en
pretty_name: Plumb (G702 curriculum study)
tags:
  - construction
  - pay-application
  - g702
  - curriculum
  - synthetic
  - verifier-as-oracle
configs:
  - config_name: benchmark
    data_files: data/benchmark.jsonl
  - config_name: train_handseeded
    data_files: data/train_handseeded.jsonl
  - config_name: train_ornith
    data_files: data/train_ornith.jsonl
  - config_name: train_blended
    data_files: data/train_blended.jsonl
  - config_name: evals
    data_files: data/evals/evals.jsonl

Plumb

Gold tasks and the held-out eval. The study is on the collection.

Config	n	What
`benchmark`	1000	Held-out eval, seed 777. Never in train.
`train_handseeded`	223	Mix matched to the eval, including PASS.
`train_ornith`	58	Ornith-1.5 proposals that passed the oracle.
`train_blended`	281	Both of the above.

Leakprobe vs benchmark: exact signature overlap 0.

curriculum	train n	sw-recall	precision	exact
hand-seeded	223	0.318 [0.290, 0.347]	0.308 [0.279, 0.337]	0.178
Ornith-only	58	0.241 [0.214, 0.268]	0.111 [0.098, 0.124]	0.084
blend	281	0.334 [0.306, 0.363]	0.374 [0.342, 0.406]	0.228

The evals config has all 15 published rows (full size, equal-N@58 random and stratified, clean equal-N@18) with their CIs, so the tables in the post are reproducible without rerunning anything:

from datasets import load_dataset

evals = load_dataset("caiotheodoro/plumb", "evals", split="train")
[r for r in evals if r["group"] == "full_size"]

Raw per-prediction dumps live at data/evals/eval-{handseeded,ornith,blended}.json and PATH_A_RESULTS.json. Those are artifacts, not configs; fetch them with hf_hub_download.

CIs: 10k bootstrap, seed 11. Adapters: handseeded · ornith · blended. N=18 mix/error and pow-* / leaked-anchor grows are not here.

from datasets import load_dataset
bench = load_dataset("caiotheodoro/plumb", "benchmark", split="train")

Apache-2.0. Synthetic pay applications, not real contractor filings.