File size: 4,852 Bytes
444a6ed
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
---
license: mit
---

# Fused Patent + arXiv Technical Clustering Dataset (Deterministic, Quality-Gated)

## Overview

This dataset is the output of a zero-touch technical clustering pipeline built over a fused corpus of patent text and arXiv-style research text.

The pipeline is fully deterministic from ingest through release and is designed to run end-to-end without manual curation or mid-run intervention. All artifacts, cluster assignments, and release decisions are derived from the same run.

This is not a curated dataset. It is a large-scale fused technical corpus that has been **deterministically analyzed and quality-gated** to isolate the portion that behaves like a semantic clustering dataset.

---

## Key Stats

* **Total labeled rows:** 9,063,272
* **Raw clusters (pre-filter):** 422
* **Release clusters (post-filter):** 147
* **Retained rows:** 3,881,329
* **Retention rate:** 42.82%
* **Shards:** 91 (labels / embeddings / chunks)
* **Size:** ~20+ GB compressed

---

## Pipeline Summary

The dataset was produced by a staged, resumable pipeline with Postgres acting as a control plane.

### Core stages

* Ingest and normalize fused patent + arXiv text
* Chunk-level embedding
* Embedding clustering
* Shard-level processing with persistent state
* Reducer-tree merge into global clusters
* Global assignment + BM25 artifact generation
* Deterministic inspection and release gating

---

## System Design

The pipeline is built to operate under real constraints (long runtimes, memory pressure, interruptions), not ideal notebook conditions.

### Control plane (Postgres)

* Task leasing and discovery
* Heartbeats and worker liveness
* Stage state tracking (not-ready / running / done / failed)
* Reducer-tree coordination and staged unblocking

### Failure-aware execution

* Distinguishes between:

  * true OOM
  * bad allocation
  * killed process
  * general memory pressure
* Descending batch ladder (deterministic step-down on failure)
* Proactive downshifting based on resource pressure
* Resumable state across interruptions

### Reducer-tree merge

* Progressive level-by-level reduction
* Final stage unblocked only after upstream completion
* Prevents global merge bottlenecks
* Avoids downstream fan-out gaps

---

## Deterministic Quality Gating

The raw clustering output was **not** treated as valid by default.

A full deterministic inspection pass across all 422 clusters produced:

* **147 coherent clusters**
* **107 mixed clusters**
* **168 metadata-heavy clusters**

### Filtering decision

For the release dataset:

* **Kept:** coherent clusters only
* **Dropped:** mixed + metadata-heavy clusters

This was done without:

* re-embedding
* hand labeling
* manual cluster editing
* modifying the original run

All decisions are reproducible from pipeline outputs.

---

## Metadata Leakage

A large portion of clusters were dominated by ingestion or wrapper fields such as:

* `source_file`
* `record_hash`
* `raw_meta_json`
* `authors_parsed`
* `published_date`
* similar structural tokens

These are not errors in the source data, but they degrade semantic clustering if left unfiltered.

Explicit detection and removal of these clusters is a core part of the release process.

---

## Dataset Structure

The release package includes filtered artifacts aligned to the retained clusters:

* `labels/` — cluster assignments
* `chunks/` — source text chunks
* `embeddings/` — embedding vectors
* `microclusters/` — original microcluster outputs (for provenance)
* `global/` — cluster summaries, BM25 artifacts, reference data

All components are consistent with the same filtered subset.

---

## What This Dataset Is

* A **deterministically derived** technical clustering dataset
* A **fused patent + research corpus** with broad technical coverage
* A **quality-gated subset** of a larger clustering run
* A **reproducible artifact** tied to a single pipeline execution

---

## What This Dataset Is Not

* Not manually curated
* Not hand-labeled
* Not cleaned via ad-hoc scripts
* Not a “perfect” semantic dataset
* Not independent from its pipeline (the pipeline defines it)

---

## Example Cluster Themes

Cluster naming was derived deterministically from top terms. Example themes include:

* wireless communication systems
* semiconductor substrates and layers
* chemical compounds and formulations
* neural networks and data processing
* vehicle control systems
* signal processing and circuits

---

## Intended Use

* Retrieval / RAG experiments
* Technical topic clustering
* Cross-domain similarity analysis
* Large-scale embedding evaluation
* Downstream filtering / refinement pipelines

---

## Notes

This dataset represents the **release-grade subset** of the full run. The original unfiltered output (422 clusters) is intentionally not presented as the primary artifact.