Datasets:

---

amazon-agi

/

RealKIE-FCC-Verified

like 3

Follow

Amazon AGI 102

Fix 14 malformed/corrupted LineItem date values in ground truth (3 documents)

Summary

While using this dataset to benchmark key-information-extraction accuracy, we
found that LineItems scores were being capped by a small number of corrupted
date strings in the ground-truth labels rather than by model behavior. This PR
corrects 14 malformed date values across 3 documents. Every correction is
verified against the source PDF image and preserves the original digit content.

No other values are changed. No document's set of rows is added to or removed
from. This is a surgical typo fix.

Documents & changes

(Row indices are 0-based within each document's LineItems array.)

To make review easier, each correction includes a bounding-box-annotated page
image (corrections/review_images/) showing where the value appears in the source document.
These were produced by OCR'ing the (scanned) pages with Amazon Textract and boxing
the corrected value using word geometry. Note: where a date repeats many times in a
spot table, the box marks a correct instance of the value rather than the exact
row — it is an orientation aid; the textual rationale in the table above is the
authoritative justification.

Methodology

1. Extracted json_response (GT) and text (OCR) for all 75 documents from the
   published parquet.
2. Programmatically flagged every LineItem*Date value that is non-null but
   either does not parse as a date or has an impossible month/day.
3. For each flagged value, rendered the source PDF at 200 DPI and visually
   confirmed the correct value and the invoice's canonical date format.
4. Applied minimal edits that preserve digit content and match the surrounding
   rows. An automated assertion verifies each "before" value matches the current
   GT exactly before writing the "after" value.

What this PR intentionally does NOT change

• Date formats. The schema types dates as free strings, so the mix of
  M-D-YY, M/D/YY, M/D/YYYY, and bare M/D is not a labeling error. We leave
  well-formed dates untouched. (Happy to provide an optional, separate
  normalization patch if the maintainers want a single canonical format.)
• Empty LineItems. The 4 documents with LineItems: [] are fee-summary
  invoices with no dated spot-detail table; [] is correct.
• Row completeness. We did not attempt to add/remove rows; that would require
  a separate, manually-verified reconciliation pass.

Attribution / license

This is a corrections contribution to a CC-BY-NC-4.0 dataset, offered under the
same terms with attribution to amazon-agi/RealKIE-FCC-Verified and the original
RealKIE dataset (arXiv 2403.20101). Source documents remain FCC public files.

Note for reviewers — the change is one Parquet file, but it's fully auditable

The only data change in this PR is data/test-00000-of-00001.parquet (the
json_response column). Since Parquet is binary, GitHub/HF won't render a diff — so
I've added a corrections/ folder to this PR branch to make the change reviewable:

• corrections/CORRECTIONS.md — the 14 edits as a table, with per-edit evidence
• corrections/MANIFEST.json — machine-readable old→new list
• corrections/gt_fixes.diff — unified diff of the affected records (pretty JSON)

Exactly 14 cells differ from main; nothing else (all values, row counts, other
columns, images, and PDFs are untouched). Every edit is a malformed/corrupted date
string corrected to the value shown in the source PDF, preserving digit content and
the document's own canonical date format.

Verify:

import pandas as pd, json
a = pd.read_parquet("hf://datasets/amazon-agi/RealKIE-FCC-Verified@main/data/test-00000-of-00001.parquet")
b = pd.read_parquet("hf://datasets/amazon-agi/RealKIE-FCC-Verified@refs/pr/3/data/test-00000-of-00001.parquet")
# diff json_response cell-by-cell -> exactly the 14 (id,row,field) in corrections/MANIFEST.json

Happy to reshape this (e.g. land the corrections as loose JSON instead of a Parquet
rewrite) if the maintainers prefer a different mechanism.

The bbox-annotated review images referenced in the PR description are now on this PR branch under corrections/review_images/ (14 PNGs, one per edit). corrections/CORRECTIONS.md links each one from its row in the corrections table. As noted there: where a date repeats in a spot table the box marks a correct instance of the value, not necessarily the exact GT row — the textual evidence column is the authoritative justification.

Cleanup: the Parquet was rewritten to preserve the original JSON key / Arrow struct-field ordering. An earlier revision of this PR had round-tripped the file through pandas, which alphabetically reordered keys in the (unchanged) json_schema column — harmless semantically, but it made the diff noisy. The PR branch now differs from main in exactly the 14 LineItems date cells listed in corrections/ and nothing else (verified cell-by-cell across all 75 rows; id, text, json_schema, and image_files are byte-identical to main).