LongListBench

GitHub | Release v2.1.0

Developed by Kay.ai.

Authors: Anton Fedoruk, Serhii Shchoholiev, and Akhil Mehta

Long-list extraction is not complete when most fields are right: an omitted, merged, or invented row can invalidate the document-level result. Across the 4 released agent runs, field micro-F1 ranges from 96.1% to 98.8%, but only 4-7 of 32 documents are complete.

LongListBench measures this failure mode in insurance and commercial trucking PDFs. A system receives one PDF or OCR transcript and a target contract, then returns the full target list. The release contains 32 synthetic PDFs and 29,599 target records; no real customer PII is included.

core_operations tests scale and output completeness. claim_multihop and policy_packets add inherited context, distant evidence, and mixed record schemas.

Complexity Stressors

Each PDF records its extraction stressors under problems:

Tag	Meaning
page_breaks	Lists or supporting evidence continue across pages with repeated headers or inherited context.
split_records	One target record has fields in separate visual blocks, sections, or pages and must be assembled.
multi_row	Records include wrapped notes, descriptions, clauses, or continuation rows.
duplicates	Prior-term, archived, duplicate, or near-duplicate distractor material is present.
large_doc	The document is long enough to stress truncation and record-completeness behavior.
multiple_tables	Target rows are mixed with summaries, ledgers, schedules, support tables, or empty tables.
multi_column	Pages use two-column or form-like layouts that stress reading order.
merged_cells	Tables include section-spanning or merged-cell structures.
ocr_condition	The released text condition is OCR from rendered page images.
ocr_layout_condition	OCR preserves visual spacing and reading order instead of converting tables into clean CSV-style rows.
long_range_evidence	Fields must be joined from distant sections of one PDF.
cross_section_join	A target record must be assembled from separately labeled sections, such as return summary, distance/gallon schedules, and liability schedules.
repeated_keys	Common keys such as states or jurisdictions repeat across sections or returns, so the key alone is insufficient for matching.
heterogeneous_record_list	A target list contains several record schemas, especially in policy packets.

These 14 tags are canonical. The manifest also retains finer audit tags, for 45 distinct problems tokens in total. Labels are metadata, not text printed inside the PDFs.

The following families make the tags easy to inspect in the PDFs:

Family or config	Stressors visible in the document
ifta_mileage_by_vehicle	Jurisdiction rows inherit unit headers while source notes interrupt page-spanning tables.
ifta_multisection_return_packet	Return headers and separate distance, fuel, and tax sections must be joined across pages; OCR preserves layout rather than clean rows.
loss_run_external	Claim rows are interleaved with descriptions, continuation notes, summaries, and empty-table distractors.
claim_multihop	Claim schedules are separated from supporting policy, driver, claimant, cause-code, and ledger sections.
policy_packets	The target mixes declarations, schedules, forms, endorsements, premiums, and clause prose.

Configs and Data Viewer

Config	Description	Target field	Documents	Records/doc range	Target records	Page range
core_operations	26 production-like commercial insurance and trucking PDFs with dense repeated operations, IFTA, and loss-run records.	records	26	260-2,571	28,178	17-84
claim_multihop	3 long claim PDFs where incident records must be assembled from distant sections.	incidents	3	12-40	77	61-148
policy_packets	3 long Businessowners, Workers Compensation, and Commercial General Liability policy packets where records must be assembled from distant sections.	records	3	344-562	1,344	99-133

Pick one config when loading:

from datasets import load_dataset

ds = load_dataset("kaydotai/LongListBench", "core_operations", split="test")
# or "claim_multihop", or "policy_packets"
print(ds)  # each row is one PDF document

Columns

Column	Type	Description
document_id	string	Stable sample identifier, e.g. ifta_mileage_by_vehicle_001 or multihop_bop_012_001.
domain	string	commercial_insurance_operations, claims, or policy_review.
complexity_regime	string	Document family, such as ifta_mileage_by_vehicle, loss_run_external, claim_crosspage_multihop, or policy_multi_hop.
evaluation_role	string	Preassigned interpretation role: scale_control or structural_challenge.
difficulty	string	Historical field retained for compatibility; in this release it stores the template or multi-hop regime.
document_format	string	Rendered layout family, currently production_like_pdf or crosspage.
num_pages	int32	Page count recorded by the generator.
target_field	string	Name of the top-level list to extract: incidents for claim multi-hop rows, records for operations, external loss-run, and policy rows.
target_record_type	string	Primary schema family, such as vehicle_state_mileage_row, driver_record, loss_run_claim_row, or policy_packet_item.
target_count	int32	Number of target records in ground_truth.
problems	list[string]	Complexity tags for the document, e.g. high_density_long_list, production_like_layout, or long_range_evidence.
transcript_conditions	list[string]	Available transcript conditions. The released rows include ocr.
pdf	Pdf	Embedded source PDF bytes.
ground_truth	string	JSON string containing the expected records under target_field.
metadata	string	JSON string with manifest metadata, file hashes, evidence maps, generation details, and source artifact paths.
ocr_transcript	string	OCR transcript generated from rendered PDF page images.

ground_truth is the complete schema-shaped object for the extraction target. Claim multi-hop rows use {"incidents": [...]}; all other list families use {"records": [...]}.

Usage

import json
from datasets import load_dataset
from datasets import Pdf

ds = load_dataset("kaydotai/LongListBench", "claim_multihop", split="test")

# The Pdf feature may decode through pdfplumber on access. Disable decoding
# when you only need bytes, transcripts, or ground truth.
ds = ds.cast_column("pdf", Pdf(decode=False))

row = ds[0]
gt = json.loads(row["ground_truth"])
records = gt[row["target_field"]]
assert len(records) == row["target_count"]

with open(f"{row['document_id']}.pdf", "wb") as f:
    f.write(row["pdf"]["bytes"])

Canonical Scoring

The tagged reference evaluator defines official scoring. Strict normalized-record completeness is primary; field overlap is a secondary diagnostic.

Method

1. Run the extractor on each PDF or transcript and return an object matching ground_truth: {"incidents": [...]} for claim multi-hop rows or {"records": [...]} for other families. The evaluator also accepts a bare list.
2. Claim incidents are keyed by normalized incident_number. Strings, dates, claimant lists, and non-zero financial breakdowns use documented canonical forms.
3. Other records normalize case, whitespace, dates, numeric formatting, accounting negatives, and documented label equivalents. Exact records are anchored before field-overlap matching; strict comparison still uses every public target field.
4. Exact-record recall is exact_record_matches / ground_truth_count. A document is complete only when the predicted and gold record multisets are identical, including duplicates and with no extra records. Record order is not scored.
5. Field recall and precision compare flattened field-value pairs; F1 is their harmonic mean. Document-macro and corpus-micro field F1 show partial correctness, not complete-list recovery.

Clone the matching release before running the example so the canonical evaluator is importable:

git clone --branch v2.1.0 --depth 1 https://github.com/kaydotai/longlistbench.git
cd longlistbench
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r benchmarks/requirements-hf.txt "pydantic>=2.5.0"

Run the scoring example from the repository root. The benchmark checkout is a source tree, not an installed Python package.

import json
from datasets import load_dataset
from benchmarks.evaluation_metrics import (
    evaluate_extraction,
    evaluate_record_extraction,
    normalize_record_predictions,
    uses_record_evaluator,
)

config = "policy_packets"
ds = load_dataset("kaydotai/LongListBench", config, split="test")

for row in ds.remove_columns("pdf"):
    gold = json.loads(row["ground_truth"])
    pred = my_predictions[row["document_id"]]
    gold_rows = gold[row["target_field"]]
    pred_rows = normalize_record_predictions(pred)
    metrics = (
        evaluate_record_extraction(pred_rows, gold_rows)
        if uses_record_evaluator(gold_rows)
        else evaluate_extraction(pred_rows, gold_rows)
    )
    print(
        row["document_id"],
        metrics["exact_record_recall"],
        metrics["complete_document"],
        metrics["f1"],
    )

Current Baselines

The release includes 4 full-corpus OCR-conditioned agentic baselines; the models and settings are listed below. Each received the OCR transcript, public field contract, and extraction prompt in a repository-denied workspace. Ground truth and target counts were unavailable.

Strict completeness on the released OCR transcripts:

Protocol	Documents	Target records	Errors	Exact-record recall	Complete documents	Field micro-F1	Field macro-F1
Codex CLI gpt-5.6-sol, xhigh reasoning	32	29,599	0	93.7%	6/32 (18.8%)	98.7%	98.3%
Claude Code CLI claude-fable-5, xhigh effort	32	29,599	0	90.9%	6/32 (18.8%)	96.1%	92.6%
Codex CLI gpt-5.5, xhigh reasoning	32	29,599	0	90.4%	4/32 (12.5%)	98.2%	97.7%
Claude Code CLI claude-opus-4-8, xhigh effort	32	29,599	0	93.6%	7/32 (21.9%)	98.8%	98.4%

The split between parser-friendly scale controls and structural challenges is visible below:

Evaluation role	Documents	Target records	GPT-5.6-Sol exact records	Fable 5 exact records	GPT-5.5 exact records	Opus 4.8 exact records
Structural challenges	19	8,414	79.0%	69.3%	67.6%	79.3%
Scale controls	13	21,185	99.5%	99.5%	99.5%	99.3%

Exact-record recall by extraction problem shows why aggregate scores should not be interpreted alone:

Extraction problem	Evaluation role	Documents	Target records	GPT-5.6-Sol exact records	Fable 5 exact records	GPT-5.5 exact records	Opus 4.8 exact records
Sparse record enrichment (driver/MVR)	Structural challenge	3	1,260	0.6%	1.9%	1.9%	1.9%
Long-range claim joins	Structural challenge	3	77	98.7%	98.7%	0.0%	98.7%
Split return schedules	Structural challenge	3	2,737	95.5%	95.5%	55.1%	95.5%
Mixed row/detail loss runs	Structural challenge	3	900	97.3%	92.2%	86.3%	97.3%
Tax inquiry detail tables	Structural challenge	2	1,300	99.9%	99.8%	100.0%	99.9%
Heterogeneous policy records	Structural challenge	3	1,344	73.3%	14.6%	95.5%	73.6%
Cross-section return joins	Structural challenge	2	796	99.6%	99.6%	99.6%	99.6%
Tax-summary scale controls	Scale control	2	1,520	99.5%	99.5%	99.5%	99.5%
Driver-schedule scale control	Scale control	1	500	99.8%	99.8%	99.8%	99.8%
Mileage-by-vehicle scale controls	Scale control	8	17,565	99.4%	99.4%	99.4%	99.2%
Vehicle-schedule scale controls	Scale control	2	1,600	100.0%	100.0%	100.0%	100.0%

Saved predictions and reports in codex_gpt56_sol_full_current_ocr_v2, claude_fable5_full_current_ocr_v2, codex_full_current_ocr_v2, claude_opus48_full_current_ocr_v2 reproduce these metrics without model access. Run metadata records the dataset hash, prompt, model settings, runtime, and output hash.

Schemas

Extraction schemas are published as standalone JSON Schema files under schemas/:

• schemas/loss_run_incident.schema.json - incidents[] for claim multi-hop incident rows, including incident identifiers, policy fields, claimant/driver fields, dates, coverage fields, and nested financial breakdowns.
• schemas/loss_run_claim_row.schema.json - records[] for external loss-run schedule rows in the core operations config.
• schemas/ifta_multisection_jurisdiction_row.schema.json - records[] for multisection IFTA return rows assembled from return headers, Schedule A mileage/gallon rows, and Jurisdictions tax-detail rows.
• schemas/policy_packet_item.schema.json - records[] for BOP/WC/CGL policy packet rows. Records are heterogeneous and may represent locations, coverages, forms, material clauses, endorsements, exclusions, rating rows, or premium items depending on record_type.

These schemas describe the strict claim, external loss-run, multisection IFTA, and policy extraction targets. Other operations rows are represented by their ground-truth field contracts and the generic record-list scorer.

Transcript Conditions

The current release includes OCR transcripts for every PDF:

• ocr_transcript: OCR text generated from 200-DPI rendered PDF page images with Google Gemini 3.5 Flash vision OCR through the direct Vertex AI API.

OCR validation reports 99.9% average identifier coverage and 99.9% tracked identifier-field support, with 17 records missing at least one tracked identifier. A separate audit finds 56 genuine OCR misses among 76,968 checked numeric fields with absolute value at least 10 (0.073%). The transcript is not hand-corrected; metadata/ocr_numeric_fidelity_baseline.json records the exact audited miss set. Interpret OCR-conditioned extraction scores with this ceiling in mind.

Provenance

The documents are synthetic. Each sample follows this generation workflow:

1. Deterministic fixtures create the schema-shaped ground truth.
2. Layout generators place those records into document-specific tables and narrative sections.
3. HTML/CSS rendering produces the source PDF.
4. OCR over rendered page images produces the released transcript for each PDF.

Policy packets follow commercial insurance structures, but all content and identifiers are synthetic.

HF rows embed the PDF, OCR transcript, ground truth, and metadata. Tagged GitHub releases also retain the rendered HTML. Private template tooling used for the current layouts is not public, so the release does not claim bit-for-bit PDF regeneration.

No customer documents or real insured, claimant, policy, or account data are included. Real documents were used only as structural references for layout and packet organization.

Limitations

LongListBench is not a substitute for evaluation on a private production corpus. Synthetic documents can underrepresent the visual and linguistic diversity of real carrier packets.

Some structured-report families are parser-friendly by design and serve as scale/completeness controls rather than hard semantic cases. Parser-transfer baselines remain useful diagnostics, but the main task allows a system to inspect each document and choose its extraction strategy. OCR support should be interpreted per affected record: one missing section header can affect many rows even when their identifiers remain visible.

License

MIT. The documents and ground truth are synthetic and released with the repository under this license.

Citation

@misc{fedoruk2026longlistbench,
  title        = {LongListBench: A Benchmark for Long-List Entity Extraction from Complex Business PDFs},
  author       = {Fedoruk, Anton and Shchoholiev, Serhii and Mehta, Akhil},
  publisher    = {Kay.ai},
  year         = {2026},
  version      = {2.1.0},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/kaydotai/LongListBench}
}