AI4Manufacturing/DD1

Multi-Modal LPBF Anomaly Detection Dataset (PB & OT)

1. Overview

This dataset is a highly structured, industry-aligned multi-modal VQA dataset designed for fine-tuning and benchmarking Vision-Language Models (VLMs) (such as Qwen2-VL, LLaVA, and CogVLM) on Laser Powder Bed Fusion (PBF-LB / LPBF) manufacturing anomaly diagnostics.

The dataset integrates two complementary physical monitoring modalities from the LPBF printing process:

• Powder Bed (PB) Images: Captured before laser melting. Excellent for detecting mechanical and geometric anomalies (e.g., blade recoating streaks, part warping, incomplete powder spreading).
• Optical Tomography (OT) Images: Captured during the laser melting phase. Essential for monitoring localized thermal accumulation, melt-pool instabilities, and keyhole defects.

By hosting both modalities in a single, unified repository with split-loading capabilities, this dataset enables researchers to train all-encompassing diagnostics models as well as perform robust ablation studies on sensor modalities.

---

2. Dataset Architecture & Schema

Every sample in this dataset is formatted using a highly standardized, enterprise-grade schema optimized for high-throughput SFT (Supervised Fine-Tuning) and secure data deduplication.

Schema Fields

Field Name	Type	Description
query	String	The system prompt/instruction directing the VLM to evaluate the printing layer.
image	Image	The serialized input image (PB or OT modality) wrapped in Apache Arrow.
annot	String	Ground-truth categorical classification label. Strictly mapped to good or anomalous.
reasoning	String	Reserved for Chain-of-Thought (CoT) explanations (currently null for base SFT).
cate	String	Category identifier. Default is "B" (indicating standard industrial SFT).
task	String	Global unique task ID. Mapped as "T-defect".
metadata	String (JSON)	Serialized metadata including modality info, original paths, and image SHA256 hashes for deduplication.
mask	Image	Reserved field for binary pixel-level defect segmentation map (currently null).
masks	List[Image]	Reserved field for multi-instance segmentation masks (currently null).

---

3. Dataset Subsets (Configs)

To support diverse experimental setups, we utilize Hugging Face's multi-configuration manager (configs) to split the dataset into three logical streams under a single repository:

Subset Breakdown

• default: A balanced, shuffled mixture of both PB and OT samples. Recommended for training unified, general-purpose LPBF QA agents.
• ot: Contains purely Optical Tomography images and corresponding anomaly annotations.
• pb: Contains purely Powder Bed visible light images and corresponding surface defect annotations.

Quick Start (How to Load)

You can flexibly stream or download any of the configurations using the Hugging Face datasets library:

from datasets import load_dataset

# 1. Load the merged multi-modal dataset (Default)
dataset_full = load_dataset("AI4Manufacturing/lpbf-structured-dataset")

# 2. Load the OT thermal emission subset only
dataset_ot = load_dataset("AI4Manufacturing/lpbf-structured-dataset", "ot")

# 3. Load the PB surface morphology subset only
dataset_pb = load_dataset("AI4Manufacturing/lpbf-structured-dataset", "pb")

# Example: Print a sample from the loaded config
print(dataset_ot["train"][0])