xll-citation-verifier — training data and evaluation

The exact 38,390-example dataset used for xll-citation-verifier: selected citation occurrences paired with three fixed decisions—recognition, usable span and cited source type.

Hub split	Role	CT examples	Federal examples	Total
train	Optimization	25,320	8,987	34,307
validation	Calibration	1,005	506	1,511
test	Internal evaluation	2,065	507	2,572
Total		28,390	10,000	38,390

Source groups/chapter families and exact full-source hashes are disjoint across the three splits. The federal tranche contains 2,000 examples from each containing source category: bills, regulations, statutory code, proposed rules and Acts.

Load the data

import json
from datasets import load_dataset
from huggingface_hub import hf_hub_download

repo = "thurstonpowers/xll-citation-verifier"
data = load_dataset(repo)
row = data["train"][0]
print(row["state"])
print(row["expected"])
questions = json.load(open(hf_hub_download(repo, "questions.json", repo_type="dataset")))
assert row["questions"] == questions

The compressed split files preserve the original native training JSONL bytes. Each row contains:

• state: exact formatted evidence supplied to Laya, including selected quote, source offsets, boundary characters and a bounded source passage.
• questions: the exact trained question definitions and answer choices.
• expected: hard targets for citation, usable_span and source_type.
• case_id, document_id, corpus, source_sha256, split_group: example, source and split identity.
• label_provenance, individually_agent_reviewed, agent_repaired, source_annotation_path: annotation provenance and original artifact reference.

Recognition/span keys are a, b, u (yes, no, uncertain). Source classes and their rules are defined in questions.json. Incomplete references can be recognition-positive and span-negative. Offsets use zero-based Python Unicode code-point indexing with an exclusive end, referring to the original source text.

Annotation and selection provenance

Labels combine user-accepted CLEF decisions with source-grounded assistant reviews and rule-assisted repairs. They are provisional reference labels. The provenance archives retain original teacher probabilities/component answers, corrected labels, repair methods and individual-review flags. The native student targets are the hard labels in expected.

CT preparation considered 28,472 candidate cases from the original 10,000-section teacher selection. The adapter admitted 28,390, retaining 55 uncertain/inconsistent annotations and 27 expanded-context holds in the exclusion ledger. The reproduction archive preserves the original 13,644-source CT inventory and its frozen 10,000 selected / 3,644 reserved partition. That original reserve is separate from the internal native splits.

Federal selection considered 12,000 candidates: 11,991 valid teacher responses and nine recorded failures. The final tranche contains 10,000 examples, with 697 holds and 1,303 eligible reserves retained in the selection ledger. The tranche includes 220 individually reviewed cases and 38 corrected label sets.

reproduction/archive-map.json maps original artifact references to their public archives and original uncompressed SHA-256 values. Historical manifests retain their original timestamps and preparation-stage status fields.

Reproduce the data and evaluation

Download an immutable snapshot, then run the standard-library verifier:

from huggingface_hub import HfApi, snapshot_download
repo = "thurstonpowers/xll-citation-verifier"
revision = HfApi().dataset_info(repo).sha
folder = snapshot_download(repo, repo_type="dataset", revision=revision)
print(folder)

python /path/to/snapshot/verify_release.py /path/to/snapshot
python /path/to/snapshot/verify_release.py /path/to/snapshot --restore-bundle ./native-bundle

Verification checks every file checksum, original split bytes, question/target identity, CT source spans, federal selected-state/label identity, source-group and hash separation, original reserve separation, preparation holds and native evaluation records. Restoration recreates the exact native training bundle.

The native exported-model evaluation contains 7,716 question items:

Question	Agree / items
Recognition	2,565 / 2,572
Usable span	2,540 / 2,572
Source type	2,549 / 2,572
Aggregate	7,654 / 7,716 (99.20%)

The separate original-reserve supplement includes the frozen 100-case source review and its matching SDK responses. The verifier reproduces recognition 100/100, span 98/100 and source-type 98/99 agreement. One source-type judgment was unresolved before scoring; the original sample contains 100 cases. These results measure agreement with provisional or assistant-reviewed references. Expert-gold legal accuracy and complete-document citation recall remain open evaluation tasks.

Training recipe

Base model: convaiinnovations/laya, revision 7b928d828b7b0e022f929d9bd2e44165aa270148, ModernBERT-large encoder. Full encoder/head fine-tuning used two epochs / 3,217 actual optimizer updates, encoder LR 2.5e-5, head LR 1e-4, effective question-item batch 64, seed 102301, and source-disjoint calibration.

code/training/ contains the exact hash-pinned training worker, thermal pacing helper and runtime Dockerfile. reproduction/ includes the original plan, runtime/start/completion and input-qualification receipts. The saved input/head budgets are 8,192/2,048 tokens, with complete native fit qualification. The training worker uses CUDA and the documented thermal-telemetry interface.

The model release identifies the evaluated weights at commit 7136187cf92d7e42277c8576486e0ca516aa483b. Apache 2.0 applies to the release annotations, code and documentation; legal source provenance is retained in the source records.