Reference Rock Databases v1.0 (1,844 specimens)

Pre-built reference databases for the rock classification system — load one and start
classifying immediately, no enrolment or index building required.

Each database contains the same 1,844 labelled rock hand-specimen images spanning
igneous, sedimentary, and metamorphic lithologies, fully indexed with ResNet-50, ResNet-18, ** VGG-16 **,VGG-19, and GoogleNet
visual features (avg_pool, 2048-D), spatial HSV colour histograms, per-specimen
textual descriptions, and text embeddings for dual-stream retrieval.

Which file should I download?

File	Descriptions written by	Text embeddings	Notes
QwenVLNet_OfflineTxtEmbd.mat	Qwen3-VL 235B	built-in offline embedding	Recommended starting point — works on any machine with zero setup
LlavaNet_OfflineTxtEmbd.mat	LLaVA	built-in offline embedding	Zero-setup alternative
QwenVLNet.mat	Qwen3-VL 235B	external embedding model	Requires the matching Ollama embedding model
LlavaNet.mat	LLaVA	external embedding model	Requires the matching Ollama embedding model

The *_OfflineTxtEmbd variants use the app's built-in deterministic embedding, so
semantic and Dual-Stream retrieval work out of the box. The other two were built with
an external Ollama embedding model; on loading, the app's embedding-consistency probe
will tell you whether your configured embedding model matches, and you can rebuild the
text embeddings from the Manage tab if it does not.

How to load

1. Launch the app (LocalOllamaRAGChat_RockClassification in MATLAB).
2. Open Manage Database → Load Database from File and select the downloaded .mat.
3. The View Database tab shows the loaded file, sample counts, and processing status.

Visual-only and Pure-CNN classification work immediately with any of the four files.
All databases were built with the single-backbone ResNet-50 pipeline; to use other
backbones or multi-CNN mode, rebuild the search index after loading (Manage tab).

Compatible with the source code at tag v1.0.0.