Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
16
16
source
stringlengths
78
113
name
stringlengths
5
20
size_bytes
int64
8
3.17k
type
stringclasses
4 values
confidence
float64
0
0.99
detect_method
stringclasses
3 values
detect_detail
stringclasses
7 values
route
stringclasses
4 values
text
stringclasses
9 values
text_chars
int64
0
237
meta
unknown
extract_ok
bool
2 classes
extract_note
stringclasses
2 values
md5
stringlengths
32
32
e9d4ef138d19a15f
D:\repos\unified-multimodal-ingestion-pipeline\data\messy_archive\archives\bundle.zip!inside/q1_report_in_zip.pdf
q1_report_in_zip.pdf
905
pdf
0.99
magic
pdf_%PDF
pdf
Quarterly ingestion report for the data platform. This document describes how mixed files are routed by type. Exact duplicates are collapsed before near duplicates are grouped. The audit log records a confidence score for every decision.
237
{ "pages": 1 }
true
91b0d3410c1d9713e3c848c16d2e25e2
d5fae3546ae775eb
D:\repos\unified-multimodal-ingestion-pipeline\data\messy_archive\archives\bundle.zip!inside/extra.txt
extra.txt
37
text
0.75
text_heuristic
decodable+ext:txt
text
a file that only exists inside a zip
36
{ "lines": 1 }
true
260731856c2f90957cf2f446ef49dd0b
0be29fa27056211d
D:\repos\unified-multimodal-ingestion-pipeline\data\messy_archive\logs\day1.txt
day1.txt
29
text
0.75
text_heuristic
decodable+ext:txt
text
line one line two line three
28
{ "lines": 3 }
true
a95cee7d8d28c9a1d6f4cd86100d341c
0541ce101aca4073
D:\repos\unified-multimodal-ingestion-pipeline\data\messy_archive\media\chart.png
chart.png
3,172
image
0.99
magic
png
image
0
{ "width": 32, "height": 32, "mode": "RGB", "format": "PNG" }
true
f9a22530b45308334265fb8d9128f56d
62f839763ad794ac
D:\repos\unified-multimodal-ingestion-pipeline\data\messy_archive\media\photo.jpg
photo.jpg
1,230
image
0.99
magic
jpeg
image
0
{ "width": 32, "height": 32, "mode": "RGB", "format": "JPEG" }
true
e0d8a87f3ab852b84c36cf698276166f
e0815fdcfd847103
D:\repos\unified-multimodal-ingestion-pipeline\data\messy_archive\notes\copy\README.md
README.md
118
text
0.75
text_heuristic
decodable+ext:md
text
# Ingestion Notes Mixed files land here in no particular order. Detection uses magic bytes first, then the extension.
118
{ "lines": 4 }
true
428e4a38f43747ee58715582fef04798
643bb99f88ecadeb
D:\repos\unified-multimodal-ingestion-pipeline\data\messy_archive\notes\README_v2.md
README_v2.md
115
text
0.75
text_heuristic
decodable+ext:md
text
# Ingestion Notes Mixed files land here in arbitrary order. Detection uses magic bytes first, then the extension.
114
{ "lines": 4 }
true
2b1e0532bcea16ec39c4800c8f292b41
29098d2085b0ac66
D:\repos\unified-multimodal-ingestion-pipeline\data\messy_archive\reports\misc\unrelated.pdf
unrelated.pdf
705
pdf
0.99
magic
pdf_%PDF
pdf
Completely different content about supply logistics. No relation to the quarterly ingestion report.
99
{ "pages": 1 }
true
144ce251069ace1b0a1be64f8953401f
5ebe5545c8b2567c
D:\repos\unified-multimodal-ingestion-pipeline\data\messy_archive\tricky\blob.bin
blob.bin
8
unknown
0
none
no_signal
unknown
0
{ "bytes": 8 }
false
unroutable
b42394dc97f5156b9e4b5e8b46cc6819
6bb98d68237fc50b
D:\repos\unified-multimodal-ingestion-pipeline\data\messy_archive\tricky\mislabeled.pdf
mislabeled.pdf
47
text
0.6
text_heuristic
decodable_no_ext
text
This is actually plain text, not a PDF at all.
46
{ "lines": 1 }
true
132ae11e39b465b879d94d21747accb4
14c683d96737ffa3
D:\repos\unified-multimodal-ingestion-pipeline\data\messy_archive\tricky\noext
noext
29
text
0.6
text_heuristic
decodable_no_ext
text
plain text with no extension
28
{ "lines": 1 }
true
8b21a0068ae9d068a322a8607db5197b

Unified Multimodal Ingestion Pipeline - flattened dataset + audit trail

This dataset is the output of the unified-multimodal-ingestion-pipeline. A synthetic messy nested archive of mixed PDF / image / text files (wrong or missing extensions, exact copies, and near-identical variants) is flattened by content type, then deduplicated in two passes (exact md5, then fuzzy MinHash/perceptual-hash), and every routing decision is recorded in a confidence-scored audit log.

  • Task category: other / data-cleaning (ingestion + deduplication)
  • License: MIT
  • Size: small representative sample (this is a demonstration corpus, not a large training set)
  • Generation method: synthetic messy archive built by the pipeline's scripts/00_make_messy_archive.py, then flattened and deduplicated by scripts/01_ingest.py. Fully reproducible on CPU.

Files

  • dataset.jsonl - one JSON record per surviving (deduplicated) document. Each record carries id, source, name, size_bytes, type, detection confidence / detect_method, extracted text, meta, and md5.
  • audit.csv - one scored decision row per input file, including detected_type, confidence, route, dedup_outcome (unique / survivor / exact_dup / near_dup), and group_survivor_id.

Measured run (shipped sample archive, CPU)

metric value
input files 17
survivors after dedup 11
exact removed (pass 1) 4
near removed (pass 2) 2
dedup rate 35.29%
unroutable files 1

Routing by detected type: text=7, pdf=5, image=4, unknown=1 (17 total).

Reproduce

python scripts/00_make_messy_archive.py
python scripts/01_ingest.py
pytest -q   # 24 passed
Downloads last month
39