thurstonpowers/xll-spo-verifier

XLL SPO Verifier — replication dataset

Dedicated replication dataset for thurstonpowers/xll-spo-verifier. Extractor data is separately maintained at thurstonpowers/xll-spo-dataset. Split from the original shared dataset without modifying any native data bytes.

Verifier splits

Hub split	Role	Assertion rows	Question items
train	Actual optimization input, including repair repetitions	58,332	174,996
validation	Source-group-reserved, deduplicated calibration	1,589	4,767
test	Original development/evaluation set	2,104	6,312

The test name makes the existing evaluation set easy to load. It is the original development/evaluation set, not newly collected or independent expert-gold data. All three roles are disjoint by the original CFR source-group split. Calibration is reserved before optimization; repair repetitions are retained in optimization and deduplicated for calibration.

The original v1.1/laya/train.jsonl contains 61,279 source-pool rows, including 34,193 unique original examples. It is not the post-calibration optimization file. Its exact SHA-256 is 23ae623356cdfb5f1fa53baa915a2b6cb5c013b3b9857fe7d428cc8bf6cb4236. The 2,104-row original dev file hashes to 3088377691c6e2d5e5217a4545ee5d26f61742a514095dca22721d8ea48a3cfa. These are the hashes bound by the published verifier's training report.

Native SDK files under v1.1/laya/native/{train,validation,test}.jsonl.gz contain only state, questions, and gold. Decompressed optimization and calibration bytes exactly match the files read back from the original training job:

• Optimization: 2c2ca9489359332e8f40e8faacae6f5721ae660671f8ce28bc13375fdcfeec71
• Calibration: a3db68d7586c9a7ffbd4362d965c7cc03a0955061a5a830188f6f0224917354d

Load the verifier data

import json
from datasets import load_dataset

data = load_dataset("thurstonpowers/xll-spo-verifier", "spo-verifier-v1_1")
row = data["train"][0]
native = {name: json.loads(row[name + "_json"])
          for name in ("state", "questions", "gold")}

The viewer stores variable-shaped native objects as JSON strings, avoiding schema inference changes to candidate fields. Every viewer row is checked for exact round-trip reconstruction. source_identifier and source_group remain visible. For native training, decompress the SDK files rather than converting viewer rows.

Questions and recipe

The frozen questions are direction_correct, claim_supported, and scope_preserved. Canonical question JSON SHA-256: d24107da8afca6260de1366b7a17e6a857b1e3a9e06b436ce9fffacfed993962. The original question file is included with the verifier reports under reproduction/xll-spo-verifier/.

The original training and early-stop finalization workers are preserved under reproduction/code/. Use Laya 0.4.2, the original pretrained convaiinnovations/laya base, sequence budget 8,192, question-head budget 2,048, seed 102301, microbatch 1, and accumulation to 64 question items. Two epochs were planned; the operator stopped after the first completed epoch. The final checkpoint was then calibrated/evaluated without more optimization. Reports preserve the actual stop and calibration history. Reproducing the recipe does not promise bitwise-identical GPU training.

Original reported evaluation: 6,312 question items, accuracy 0.949, ECE 0.019, Brier 0.091. These are preserved original metrics, not a fresh full model reevaluation by this dataset release. The fitted noul temperature reached the original clamp, and the runtime has documented out-of-range temperature fallbacks; retain the model card's calibration caveat.

Verify the release

Download a pinned dataset revision and run:

python verify_release.py /path/to/downloaded/dataset

The verifier checks all release hashes, native split hashes/counts, viewer round-trips, source-group disjointness and the model-bound reports. Consult reproduction/public-release.json for the exact model and dataset parent revisions.

Source and label provenance

Source text is public-domain federal eCFR/GPO material. The v1.1 cohort considered 5,000 federal sections, accepted 4,960 and retained 40 unresolved exclusions. Labels derive from teacher extraction/verification and subsequent agent review and repair. They are not independent expert legal gold. Connecticut was excluded from these v1.1 training exports. Original sampling, provenance, unresolved counts and data hashes remain in v1.1/manifest.json and the source rows.

Historical version boundary

The original shared repository retains the historical provision-oriented root files and original verifier copies for provenance and compatibility. Use this dedicated repository for current verifier replication. No legacy root data is included here. Labels and original metrics have not been freshly reevaluated.

License

Apache-2.0 for the dataset packaging and released annotations; federal source text is a U.S. government work. Model/base-model licenses apply separately.