Datasets:
id stringlengths 11 11 | source_identifier stringlengths 15 28 | source_group stringlengths 13 20 | row_json stringlengths 9.22k 93.6k |
|---|---|---|---|
train-00000 | /us/cfr/t1/s19.5 | /us/cfr/t1/p19 | {"document_id":"54df9123e75a4da8610bd0900cf6c241128fce3d7925971bb707447ad440b7a0","source_identifier":"/us/cfr/t1/s19.5","source_version":"2026-09-17","source_sha256":"6a69751c852c87756d31ea06ead573ef2d2c4f5dead79628b6ece64f167a3dcc","split":"train_candidate","split_group":"/us/cfr/t1/p19","prior_training_overlap":true... |
train-00001 | /us/cfr/t1/s19.5 | /us/cfr/t1/p19 | {"document_id":"54df9123e75a4da8610bd0900cf6c241128fce3d7925971bb707447ad440b7a0","source_identifier":"/us/cfr/t1/s19.5","source_version":"2026-09-17","source_sha256":"6a69751c852c87756d31ea06ead573ef2d2c4f5dead79628b6ece64f167a3dcc","split":"train_candidate","split_group":"/us/cfr/t1/p19","prior_training_overlap":true... |
train-00002 | /us/cfr/t1/s19.5 | /us/cfr/t1/p19 | {"document_id":"54df9123e75a4da8610bd0900cf6c241128fce3d7925971bb707447ad440b7a0","source_identifier":"/us/cfr/t1/s19.5","source_version":"2026-09-17","source_sha256":"6a69751c852c87756d31ea06ead573ef2d2c4f5dead79628b6ece64f167a3dcc","split":"train_candidate","split_group":"/us/cfr/t1/p19","prior_training_overlap":true... |
train-00003 | /us/cfr/t1/s8.9 | /us/cfr/t1/p8 | {"document_id":"6435006919ea1494a18fcc7ba67d11646177a6f310806ba367f6528c14215d16","source_identifier":"/us/cfr/t1/s8.9","source_version":"2026-09-17","source_sha256":"0d15fdac333da5454304640d7f2d154d5db86641c0b5d3c84be0bde5d6e51049","split":"train_candidate","split_group":"/us/cfr/t1/p8","prior_training_overlap":true,"... |
train-00004 | /us/cfr/t36/s214.18 | /us/cfr/t36/p214 | {"document_id":"9ce54c235d095058906c290de428a673deb8132e7a7e216b5ae111c30a147e11","source_identifier":"/us/cfr/t36/s214.18","source_version":"2026-09-17","source_sha256":"e16953f8b6f8456cd16dfa48d386d730704ed71cde4c492f36ecafb4fb6e7c81","split":"train_candidate","split_group":"/us/cfr/t36/p214","prior_training_overlap"... |
train-00005 | /us/cfr/t36/s214.18 | /us/cfr/t36/p214 | {"document_id":"9ce54c235d095058906c290de428a673deb8132e7a7e216b5ae111c30a147e11","source_identifier":"/us/cfr/t36/s214.18","source_version":"2026-09-17","source_sha256":"e16953f8b6f8456cd16dfa48d386d730704ed71cde4c492f36ecafb4fb6e7c81","split":"train_candidate","split_group":"/us/cfr/t36/p214","prior_training_overlap"... |
train-00006 | /us/cfr/t36/s214.18 | /us/cfr/t36/p214 | {"document_id":"9ce54c235d095058906c290de428a673deb8132e7a7e216b5ae111c30a147e11","source_identifier":"/us/cfr/t36/s214.18","source_version":"2026-09-17","source_sha256":"e16953f8b6f8456cd16dfa48d386d730704ed71cde4c492f36ecafb4fb6e7c81","split":"train_candidate","split_group":"/us/cfr/t36/p214","prior_training_overlap"... |
train-00007 | /us/cfr/t25/s37.102 | /us/cfr/t25/p37 | {"document_id":"2c4ca3647169b7cc1aa2d8cbc56a0eaef1aee4a350ed4d125e10fcb586590519","source_identifier":"/us/cfr/t25/s37.102","source_version":"2026-09-17","source_sha256":"ae2bf37d0878c6552a5e6091bcfc1c8d639960042405385a4dcfda70c683cd32","split":"train_candidate","split_group":"/us/cfr/t25/p37","prior_training_overlap":... |
train-00008 | /us/cfr/t25/s37.102 | /us/cfr/t25/p37 | {"document_id":"2c4ca3647169b7cc1aa2d8cbc56a0eaef1aee4a350ed4d125e10fcb586590519","source_identifier":"/us/cfr/t25/s37.102","source_version":"2026-09-17","source_sha256":"ae2bf37d0878c6552a5e6091bcfc1c8d639960042405385a4dcfda70c683cd32","split":"train_candidate","split_group":"/us/cfr/t25/p37","prior_training_overlap":... |
train-00009 | /us/cfr/t25/s37.102 | /us/cfr/t25/p37 | {"document_id":"2c4ca3647169b7cc1aa2d8cbc56a0eaef1aee4a350ed4d125e10fcb586590519","source_identifier":"/us/cfr/t25/s37.102","source_version":"2026-09-17","source_sha256":"ae2bf37d0878c6552a5e6091bcfc1c8d639960042405385a4dcfda70c683cd32","split":"train_candidate","split_group":"/us/cfr/t25/p37","prior_training_overlap":... |
train-00010 | /us/cfr/t28/s124.22 | /us/cfr/t28/p124 | {"document_id":"6028f37a7a86a71cd2b84e62781d67630e8bae448209b96a34693b9b5982feac","source_identifier":"/us/cfr/t28/s124.22","source_version":"2026-09-17","source_sha256":"4196e8afb63b0b706dced69ce9f8017f3435cf35432cf4fadd1c16ef3f100fe9","split":"train_candidate","split_group":"/us/cfr/t28/p124","prior_training_overlap"... |
train-00011 | /us/cfr/t36/s9.180 | /us/cfr/t36/p9 | {"document_id":"54da00761487e582c81def19a02f8118bff4f87cd274dd1aef3251f44531b7d5","source_identifier":"/us/cfr/t36/s9.180","source_version":"2026-09-17","source_sha256":"2e381a4d0da95e29f663ee0e89d58ed0ff3ecc02952d7c71c55b0bd6fb23d446","split":"train_candidate","split_group":"/us/cfr/t36/p9","prior_training_overlap":fa... |
train-00012 | /us/cfr/t41/s51-9.503 | /us/cfr/t41/p51-9 | {"document_id":"c74c4e8a8576ce207a07f95b30f31d19fa2994e080395891f1e16d80389786e1","source_identifier":"/us/cfr/t41/s51-9.503","source_version":"2026-09-17","source_sha256":"d34bf6fde4a71bdad832bfb5ad4af52fe30e7f92e0da2f806f004e9dca639b82","split":"train_candidate","split_group":"/us/cfr/t41/p51-9","prior_training_overl... |
train-00013 | /us/cfr/t33/s386.7 | /us/cfr/t33/p386 | {"document_id":"c5afa3f412f623679d0319d487c94cba25a3de9744e0d8ecdaedf956160e478f","source_identifier":"/us/cfr/t33/s386.7","source_version":"2026-09-17","source_sha256":"a3d0074735e96a93f0e359cc55f46c310e6719f106bef1cfd364de57fcda951b","split":"train_candidate","split_group":"/us/cfr/t33/p386","prior_training_overlap":... |
train-00014 | /us/cfr/t33/s386.7 | /us/cfr/t33/p386 | {"document_id":"c5afa3f412f623679d0319d487c94cba25a3de9744e0d8ecdaedf956160e478f","source_identifier":"/us/cfr/t33/s386.7","source_version":"2026-09-17","source_sha256":"a3d0074735e96a93f0e359cc55f46c310e6719f106bef1cfd364de57fcda951b","split":"train_candidate","split_group":"/us/cfr/t33/p386","prior_training_overlap":... |
train-00015 | /us/cfr/t33/s386.7 | /us/cfr/t33/p386 | {"document_id":"c5afa3f412f623679d0319d487c94cba25a3de9744e0d8ecdaedf956160e478f","source_identifier":"/us/cfr/t33/s386.7","source_version":"2026-09-17","source_sha256":"a3d0074735e96a93f0e359cc55f46c310e6719f106bef1cfd364de57fcda951b","split":"train_candidate","split_group":"/us/cfr/t33/p386","prior_training_overlap":... |
XLL SPO Extractor — replication dataset
Dedicated model datasets
This is the canonical dataset for xll-spo-nemotron-4b actor extraction.
The verifier now has its own dataset: xll-spo-verifier.
Original verifier copies below remain frozen for provenance and existing downloads;
use the dedicated verifier repository for new work. The default configuration here
is spo-extractor-v1_1.
Exact source data for:
- xll-spo-verifier: three-question actor-relation verification, one completed optimization epoch followed by calibration and evaluation.
- xll-spo-nemotron-4b: actor → relationship → target extraction, two completed epochs.
This dataset was previously private. The public release preserves the original training files and adds native SDK splits, stable viewer configurations, training code, model-bound reports, hashes and a standalone reproduction check.
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-dataset", "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.
Extractor data
v1.1/nemotron/train.jsonl: 7,462 chat-format examples, 4,660 unique,
with repaired targets repeated three times. dev.jsonl: 300 reviewed
CFR-part-group development examples. The published extractor binds the training
file's SHA-256:
b27b4bd55329547599254954219092e02c137fa19f5404204387dc05e7b6a98a.
Load configuration spo-extractor-v1_1; deserialize row_json for the unchanged
original record, or download the original JSONL files directly.
Its original greedy evaluation report and evaluator are included. Reported SPO surface-agreement F1 0.4237 and qualifier/participant F1 0.3025 use the 300-section development denominator, retaining failed/capped sections. These are reference-agreement metrics, not semantic correctness or complete legal coverage.
Verify the release
Download a pinned dataset revision and run:
python verify_release.py /path/to/downloaded/dataset
The verifier checks original file 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
Root-level train.jsonl, verifier_train.jsonl, verifier_dev.jsonl, prompts,
governance ledgers and evaluation reports belong to the older provision-oriented
v3 contract. They are preserved as history, not inputs for reproducing the current
actor models. The previous dataset card is retained at
reproduction/README.before-public-20261109.md. The two contracts are distinct.
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.
- Downloads last month
- 75