Arch-300K: Dual-Use Corpora in Architecture & Autoscaling

• Arch-300K is a dual-use label representing a multimodal architectural VQA dataset of 315K image–question–answer triplets and a weakly supervised workload window corpus from Azure Functions traces.
• The architectural corpus uses a multi-stage pipeline combining image filtering, LLM-guided text verification, and aspect-driven annotation to enhance visually grounded architectural analysis.
• The workload corpus leverages sliding-window feature extraction and weak supervision with labeling functions to classify serverless workloads for uncertainty-aware autoscaling.

Arch-300K denotes two unrelated research corpora that share the same label in 2025 arXiv literature. In "ArchGPT: Understanding the World's Architectures with Large Multimodal Models," Arch-300K is a domain-specialized architectural visual question answering corpus of 315,247 image–question–answer triplets, built from Wikimedia Commons scenes and used to fine-tune ArchGPT for visually grounded architectural analysis (Wang et al., 25 Sep 2025). In "Archetype-Aware Predictive Autoscaling with Uncertainty Quantification for Serverless Workloads on Kubernetes," "Arch-300K" is an accurate shorthand for a weakly supervised corpus of 300,000+ workload windows derived from Azure Functions traces and labeled into four workload archetypes for AAPA; the paper does not introduce it as a formal dataset name (Zhang et al., 8 Jul 2025).

1. Terminological scope and disambiguation

The shared label masks a substantive domain difference. One Arch-300K concerns multimodal architectural interpretation, with image-question-answer supervision, scene-level curation, and long-form annotations. The other concerns serverless workload archetyping, with sliding windows over invocation traces, feature extraction, labeling functions, and autoscaling simulation. The overlap is nominal rather than methodological.

Context	Meaning of Arch-300K	Status
ArchGPT	A domain-specialized dataset of approximately 315,000 image-question-answer triplets centered exclusively on architecture	Explicit dataset name
AAPA	A weakly supervised corpus of over 300,000 workload windows constructed from public Azure Functions traces and labeled into four time-series archetypes	Accurate shorthand, not a formal dataset name

This distinction matters for bibliographic precision. In the ArchGPT work, Arch-300K is the central data asset around which the model, pipeline, and evaluation are organized. In the AAPA work, the label serves as a convenient identifier for a reconstructed corpus that underpins weak supervision, classification, and autoscaling evaluation (Wang et al., 25 Sep 2025, Zhang et al., 8 Jul 2025).

2. Arch-300K in architectural multimodal learning

Within the ArchGPT framework, Arch-300K is defined as a curated set of approximately 315,000 image–question–answer items centered exclusively on architecture, with two complementary annotation modes: detailed descriptions and aspect-guided conversations. It was created to address a persistent gap: general-purpose LMMs trained on internet-scale image–text corpora tend to hallucinate and rely on language priors when asked about architectural style, elements, materials, symbolic motifs, or historical context. The dataset is intended to inject expert-style, visually grounded supervision across thousands of architectural scenes, reducing hallucinations and strengthening part-, style-, and material-level reasoning (Wang et al., 25 Sep 2025).

Its source domain is Wikimedia Commons categories treated as scenes, reusing the MegaScenes protocol and crawl results, where Commons categories are linked to Wikidata items with multilingual, structured metadata. Each scene aggregates unconstrained tourist photo collections uploaded over time. Architectural scenes are identified by a keyword dictionary plus an LLM gate: scene names or taxonomies containing terms in an architecture dictionary, such as building, arch, or chapel, are accepted; those matching non-architectural terms such as nature or activity are rejected; ambiguous cases are resolved by Gemini 2.5 Pro via the binary prompt “Is $an architectural structure?” Yes/No. About 40% of Commons scenes are architectural at this stage. The dataset’s geographic, stylistic, and typological coverage is global. The supplementary material lists examples including church, monument, palace, lighthouse, rathaus/town hall, mosque, temple, and synagogue, spanning varied regions, historical eras, functions, and stylistic lineages. Images reflect diverse viewpoints, cameras, seasons, lighting, and weather typical of tourist photo collections. The dataset is explicitly designed to support VR/MR/AR scenarios, including interactive guided tours, ask-as-you-explore VR walkthroughs, and AR overlays that answer questions from arbitrary viewpoints while citing visible cues rather than generic priors. ## 3. Construction pipeline and annotation schema in the architectural corpus The ArchGPT paper describes Arch-300K as a multi-stage pipeline that cleans visual inputs and distills reliable domain text scene-by-scene. The first stage is coarse-to-fine image filtering integrating 3D reconstruction and segmentation. For each scene with unconstrained images \(I_1,\ldots,I_N\), capped at 200 to avoid OOM, VGGT estimates a scene point cloud \(P\) and per-image confidence maps \(C_1,\ldots,C_N\). Sky regions are removed via ONNX sky segmentation. A global threshold \(\tau\) is computed from the \(\alpha\)-quantile of confidence values with \(\alpha = 0.8\), and each confidence map is thresholded to produce a coarse binary mask \(c_i(x,y)\in\{0,1\}\), where \(c_i(x,y)=1\) if \(C_i(x,y)\ge \tau\), else \(0\). The per-image valid-coverage ratio is$ c_i = \frac{1}{H \times W} \sum_{x=1}^{H}\sum_{y=1}^{W} c_i(x,y).

$Images with \(c_i\) below a fixed cutoff are discarded. For images that pass coarse filtering, \(P=10\) positive point prompts are sampled uniformly from \(c_i=1\) regions and passed to SAM with a ViT-H backbone to produce refined masks \(f_i\). The fine valid-pixel ratio is$

f_i = \frac{1}{H \times W} \sum_{x=1}^{H}\sum_{y=1}^{W} f_i(x,y).

$Images are ranked by \(f_i\), and the final architectural frames are selected by$

{I'1,\ldots,I'_K} = \mathrm{TopK}({f_i}{i=1}^{N}, K), $with \(K=8\). Negative prompts were found unhelpful and omitted. Around 55% of architectural scenes are entirely dropped at this step because no image survives filtering. The second stage is LLM-guided text verification and knowledge distillation. A Gemini 2.5 Pro “LLM-known checker” verifies whether the scene is within the model’s knowledge scope using the prompt “You are a professional architect. Do you know the$? Respond with only ‘YES’ or ‘NO’.” Scenes answered NO are discarded. A second Gemini 2.5 Pro step, described as a Wiki-data extractor and used in a RAG-like manner, extracts three fields strictly from provided Wikimedia/Wikidata text: a refined architectural description, the geographic location and construction year, and the formal architecture name. If a detail is absent, the output is set to “Unknown.” Prompts explicitly instruct “DO NOT incorporate your own knowledge; ONLY use provided material.” The paper reports that location and year extraction are more accurate when restricted to Wikidata-like structured text.

The third stage synthesizes formal analysis annotations. In the detailed-description mode, paragraph-length prompts modeled on LLaVA-style instruction formats are conditioned on formal name, location, year, and refined description, but the LLM is instructed to focus only on visible characteristics and to avoid naming the building, beginning with “In this image, I can see...” or “This image shows...”. In the aspect-guided conversation mode, the LLM selects up to five aspects from a predefined taxonomy—architectural style, architectural elements, architectural details, architectural context, architectural innovation, architectural symbolism, architectural materials—and may add others if relevant. For each selected aspect, it generates a concise question of 10–15 words beginning with “What/How/Which/Describe...” and a long-form answer beginning with “In this image, the {aspect}...” or “This image shows that the {aspect}...”, again avoiding the building’s name. Final inclusion is preceded by manual screening and refinement. Ablation results indicate substantial gains from retaining both the LLM-known checker and the Wiki-data extractor (Wang et al., 25 Sep 2025).

4. Composition, training role, and empirical behavior of the architectural corpus

The architectural Arch-300K contains 315,247 VQA pairs drawn from 8,643 unique architectural scenes, with a maximum of 8 images per scene and an average of 72.9 questions per scene. The reported split is 263,806 training items and 51,441 testing items, with no explicit validation split. Its annotation modes are distributed as 23% detailed descriptions and 77% aspect-guided conversations. Average answers are 81.7 words, and the corpus emphasizes open-ended, long-form generation rather than short answers, binary yes/no, or template counting questions. Annotations are constructed in English, although evaluation reports multilingual robustness in English and Chinese despite the absence of multilingual training annotations (Wang et al., 25 Sep 2025).

The dataset is used to supervise ArchGPT by fine-tuning ShareGPT4V-7B following the LLaVA paradigm. The vision encoder is CLIP ViT-L/14 at $336\times336$, producing 576 visual tokens. A two-layer MLP projector maps visual tokens into the LLM embedding space. The LLM is Vicuna-v1.5 7B, LLaMA2-based, adapted via LoRA with rank $r=128$ and $\alpha=256$; the base LLM and vision tower are frozen. The training objective is standard next-token likelihood on answers with a canonical VQA cross-entropy loss

$L_{\mathrm{VQA}} = -\sum_t \log p(a_t \mid I, q),$

and a total loss

$L = L_{\mathrm{VQA}} + \lambda L_{\mathrm{aux}},$

where optional auxiliary terms are not specified. Optimization uses batch size 16, learning rate $2\times10^{-5}$, two epochs, AdamW with cosine schedule and 3% warmup, weight decay 0.0, max context length 2048, bfloat16 with TF32 matmuls, gradient checkpointing, and DeepSpeed ZeRO-3 on a single 40 GB GPU. Multi-annotation fusion avoids redundant image encodes and cuts training time by more than 50% with negligible sequence overhead.

Evaluation on Arch-300K uses JudgeLM scoring under both vanilla and architecture-tailored prompts, together with GLEU, METEOR, and ROUGE-1/2/L. ArchGPT obtains $Images with \(c_i\) below a fixed cutoff are discarded. For images that pass coarse filtering, \(P=10\) positive point prompts are sampled uniformly from \(c_i=1\) regions and passed to SAM with a ViT-H backbone to produce refined masks \(f_i\). The fine valid-pixel ratio is$0, compared with InternVL3-8B at 7.150, LLaVA-1.5-7B at 7.472, ShareGPT4V-7B at 6.348, and Qwen-VL-Chat-7B at 6.182. Its architecture-focused $Images with \(c_i\) below a fixed cutoff are discarded. For images that pass coarse filtering, \(P=10\) positive point prompts are sampled uniformly from \(c_i=1\) regions and passed to SAM with a ViT-H backbone to produce refined masks \(f_i\). The fine valid-pixel ratio is$1, compared with 6.554, 6.779, 5.624, and 5.490 respectively. It also leads in GLEU 14.43, METEOR 30.45, ROUGE-1 40.42, ROUGE-2 12.72, and ROUGE-L 24.91. JudgeLM aspect-wise preference assigns it the highest win rates across creativity, level of detail, logical consistency, and domain expertise. Ablations show that removing coarse/fine filtering, the LLM-known checker, or the Wiki-data extractor reduces performance; the full pipeline raises $Images with \(c_i\) below a fixed cutoff are discarded. For images that pass coarse filtering, \(P=10\) positive point prompts are sampled uniformly from \(c_i=1\) regions and passed to SAM with a ViT-H backbone to produce refined masks \(f_i\). The fine valid-pixel ratio is$2 from 6.079 to 7.713 and $Images with \(c_i\) below a fixed cutoff are discarded. For images that pass coarse filtering, \(P=10\) positive point prompts are sampled uniformly from \(c_i=1\) regions and passed to SAM with a ViT-H backbone to produce refined masks \(f_i\). The fine valid-pixel ratio is$3 from 5.234 to 7.107. The model largely preserves performance on general multimodal benchmarks including MMB, LLaVA-W, MM-Vet, and ScienceQA, and the paper reports no catastrophic forgetting.

5. Arch-300K as a workload-window corpus for AAPA

In the AAPA paper, Arch-300K refers to a weakly supervised corpus of over 300,000 workload windows built from public Azure Functions traces and labeled into four time-series archetypes: PERIODIC, SPIKE, RAMP, and STATIONARY_NOISY. The corpus exists to enable scalable, automatic workload archetyping and archetype-aware autoscaling without manual labeling. Its provenance is the public Azure Functions dataset characterized by Shahrad et al. at USENIX ATC 2020, which the paper notes reveals invocation rates spanning 8+ orders of magnitude. The authors do not redistribute the traces or state a dataset license; instead, they open-source simulator and analysis code, including feature extraction logic and labeling functions, and users reproduce the corpus by obtaining the cited traces and running the provided pipelines (Zhang et al., 8 Jul 2025).

The dataset is formed by sliding a 60-minute window with 10-minute stride over HTTP-triggered functions with at least 1,000 total invocations. Each window is represented by 38 features spanning statistical, time-domain, and frequency-domain properties, including mean, variance, kurtosis, max-to-median ratio, autocorrelation metrics, spectral entropy, and linear-fit trend features. The paper reports two distributions. In the preprocessing section, the natural distribution is approximately 35% SPIKE, 30% STATIONARY_NOISY, 25% PERIODIC, and 10% RAMP. In the classification section, the labeled train/test splits are heavily imbalanced—PERIODIC 70.2%, SPIKE 17.6%, STATIONARY_NOISY 12.0%, and RAMP 0.2%—and inverse-frequency class weights are used.

Weak supervision is implemented through ten domain-specific labeling functions with majority voting. Representative criteria include

$Images with \(c_i\) below a fixed cutoff are discarded. For images that pass coarse filtering, \(P=10\) positive point prompts are sampled uniformly from \(c_i=1\) regions and passed to SAM with a ViT-H backbone to produce refined masks \(f_i\). The fine valid-pixel ratio is$4

for bursty windows, and

$Images with \(c_i\) below a fixed cutoff are discarded. For images that pass coarse filtering, \(P=10\) positive point prompts are sampled uniformly from \(c_i=1\) regions and passed to SAM with a ViT-H backbone to produce refined masks \(f_i\). The fine valid-pixel ratio is$5

for periodic windows, where $Images with \(c_i\) below a fixed cutoff are discarded. For images that pass coarse filtering, \(P=10\) positive point prompts are sampled uniformly from \(c_i=1\) regions and passed to SAM with a ViT-H backbone to produce refined masks \(f_i\). The fine valid-pixel ratio is$6 is spectral entropy and $Images with \(c_i\) below a fixed cutoff are discarded. For images that pass coarse filtering, \(P=10\) positive point prompts are sampled uniformly from \(c_i=1\) regions and passed to SAM with a ViT-H backbone to produce refined masks \(f_i\). The fine valid-pixel ratio is$7 is autocorrelation at the dominant period. RAMP windows are identified by strong linear trends, and STATIONARY_NOISY windows by near-stationary behavior with random fluctuations while failing spike and periodicity checks. Let $Images with \(c_i\) below a fixed cutoff are discarded. For images that pass coarse filtering, \(P=10\) positive point prompts are sampled uniformly from \(c_i=1\) regions and passed to SAM with a ViT-H backbone to produce refined masks \(f_i\). The fine valid-pixel ratio is$8 denote the number of labeling functions and $Images with \(c_i\) below a fixed cutoff are discarded. For images that pass coarse filtering, \(P=10\) positive point prompts are sampled uniformly from \(c_i=1\) regions and passed to SAM with a ViT-H backbone to produce refined masks \(f_i\). The fine valid-pixel ratio is$9. Each labeling function $Images are ranked by \(f_i\), and the final architectural frames are selected by$0 outputs $Images are ranked by \(f_i\), and the final architectural frames are selected by$1, and the aggregated label is

$Images are ranked by \(f_i\), and the final architectural frames are selected by$2

Confidence is proportional to agreement:

$Images are ranked by \(f_i\), and the final architectural frames are selected by$3

The resulting corpus trains a LightGBM classifier, with per-class probabilities calibrated via beta calibration. The calibrated confidence drives uncertainty-aware autoscaling. AAPA uses per-archetype strategies—warm pools for SPIKE, predictive pre-scaling for PERIODIC, trend extrapolation for RAMP, and conservative scaling for STATIONARY_NOISY—and adjusts aggressiveness with

$Images are ranked by \(f_i\), and the final architectural frames are selected by$4

$Images are ranked by \(f_i\), and the final architectural frames are selected by$5

$Images are ranked by \(f_i\), and the final architectural frames are selected by$6

$Images are ranked by \(f_i\), and the final architectural frames are selected by$7

Replica decisions follow

$Images are ranked by \(f_i\), and the final architectural frames are selected by$8

with cooldown-bounded control, and for PERIODIC workloads a proactive forecast floor

$Images are ranked by \(f_i\), and the final architectural frames are selected by$9

Empirically, the LightGBM classifier achieves 99.8% accuracy on the test set. In autoscaling evaluation, AAPA reduces SLO violations by up to 50%, with a reported example of 1.8% versus 3.6% for HPA on STATIONARY_NOISY workloads, and improves response time by about 40%, while resource usage rises by 2–8$with \(K=8\). Negative prompts were found unhelpful and omitted. Around 55% of architectural scenes are entirely dropped at this step because no image survives filtering. The second stage is LLM-guided text verification and knowledge distillation. A Gemini 2.5 Pro “LLM-known checker” verifies whether the scene is within the model’s knowledge scope using the prompt “You are a professional architect. Do you know the$0 the pod-minutes of HPA, with SPIKE-heavy workloads showing the largest overhead. The baselines are Kubernetes HPA with target CPU 70% and stabilization 5 minutes, and a Generic Predictive baseline using uniform Holt–Winters with a 15-minute horizon. Simulation uses 30-second pod startup, 1-minute metric aggregation, FIFO queueing, max replicas 100, CPU 1000 millicores per replica, and 256 MB memory per replica. The paper also defines a Resource Efficiency Index,

$with \(K=8\). Negative prompts were found unhelpful and omitted. Around 55% of architectural scenes are entirely dropped at this step because no image survives filtering. The second stage is LLM-guided text verification and knowledge distillation. A Gemini 2.5 Pro “LLM-known checker” verifies whether the scene is within the model’s knowledge scope using the prompt “You are a professional architect. Do you know the$1

with $with \(K=8\). Negative prompts were found unhelpful and omitted. Around 55% of architectural scenes are entirely dropped at this step because no image survives filtering. The second stage is LLM-guided text verification and knowledge distillation. A Gemini 2.5 Pro “LLM-known checker” verifies whether the scene is within the model’s knowledge scope using the prompt “You are a professional architect. Do you know the$2, $with \(K=8\). Negative prompts were found unhelpful and omitted. Around 55% of architectural scenes are entirely dropped at this step because no image survives filtering. The second stage is LLM-guided text verification and knowledge distillation. A Gemini 2.5 Pro “LLM-known checker” verifies whether the scene is within the model’s knowledge scope using the prompt “You are a professional architect. Do you know the$3, and $with \(K=8\). Negative prompts were found unhelpful and omitted. Around 55% of architectural scenes are entirely dropped at this step because no image survives filtering. The second stage is LLM-guided text verification and knowledge distillation. A Gemini 2.5 Pro “LLM-known checker” verifies whether the scene is within the model’s knowledge scope using the prompt “You are a professional architect. Do you know the$4 by default.

6. Limitations, bias, and scholarly significance

The architectural Arch-300K inherits the strengths and liabilities of its construction choices. Its images are sourced from Wikimedia Commons, whose files carry original licenses and structured metadata; image licensing follows Commons’ per-file terms, while the specific annotation license is not stated. Privacy exposure is reduced because the pipeline favors occlusion-free and architecture-focused frames and removes people and cars when possible. The paper nevertheless identifies biases and limitations: reliance on LLM knowledge priors, underrepresentation of lesser-known or sparsely documented buildings, “Unknown” fields when facts are missing, tourist-photo biases toward iconic sites and daylight conditions, and a scale that is sufficient for fine-tuning 7B-scale LMMs but modest for much larger models. The AAPA corpus has a different failure surface: majority-vote labeling functions may mislabel edge cases if thresholds are not retuned, archetype boundaries may drift as applications evolve, the rare RAMP class remains hard, reported class distributions vary across sections, and the simulator abstracts networking and contention; the paper also notes a minor inconsistency in startup times between simulator configuration and discussion (Wang et al., 25 Sep 2025, Zhang et al., 8 Jul 2025).

A common misconception is that Arch-300K uniquely names the architectural dataset. The AAPA paper explicitly supports a second usage, but only as an accurate shorthand rather than a formal dataset title. The two corpora are methodologically orthogonal. One operationalizes architectural formal analysis through 3D/segmentation-based image curation, Wikidata-grounded text normalization, and long-form aspect taxonomy. The other operationalizes workload archetyping through sliding-window feature extraction, weak supervision, LightGBM classification, calibrated uncertainty, and autoscaling simulation. A plausible implication is that scholarly disambiguation requires pairing the label with its surrounding framework—ArchGPT for multimodal architecture, AAPA for serverless autoscaling—rather than treating “Arch-300K” as a uniquely identifying dataset name.