How teba-cs/cunei-restore works
CuneiML v3 is a multimodal Cuneiform restoration and dataset project: it acquires linked tablet images, builds a validated text-and-asset dataset, trains and evaluates restoration models, and produces uncertainty-aware line and tablet restorations. The graph follows the documented and sampled workflow; several subsystem links rely on module names and documented artifacts because their implementation was not sampled.
Main entry points: build_parquet_v3.py Linked dataset · prepare_examples_v3.py Restoration examples · deploy_automatic_restoration_v3.py Automatic restoration · download_images.py Image downloader
Components 20
Data preparation
- Cuneiform source
CuneiMLv1.2_multimodal_dataset_v3.py - Source audit
audit_source_v3.py - Image downloader
download_images.py - Download state
download_images.py - Linked dataset
build_parquet_v3.py - Frozen splits
freeze_splits_v3.py - Restoration examples
prepare_examples_v3.py
Restoration models
- Visual embeddings
extract_visual_embeddings_v3.py - Model training
train_five_models_v3.py - Architecture search
train_m0_search_v3.py
Evaluation
- Model selection
evaluate_selection_v3.py - Confidence calibration
predict_calibration_v3.py - Final model lock
calibrate_and_lock_v3.py - Protected evaluation
evaluate_protected_test_v3.py - Results analysis
analyze_and_report_v3.py
Restoration delivery
- Restoration runtime
runtime_v3.py - Automatic restoration
deploy_automatic_restoration_v3.py - Line and tablet results
outputs
Other components
- Researcher
- CDLI image host
Connections 25
- Researcher → Source audit: audits source
- Cuneiform source → Source audit: validates records
- Researcher → Image downloader: requests assets
- Image downloader → CDLI image host: downloads images
- Image downloader → Download state: records state
- Source audit → Linked dataset: provides audited records
- Download state → Linked dataset: supplies asset metadata
- Linked dataset → Frozen splits: provides tablet data
- Linked dataset → Restoration examples: provides lines and assets
- Frozen splits → Restoration examples: gates example creation
- Linked dataset → Visual embeddings: provides images
- Restoration examples → Model training: provides training data
- Visual embeddings → Model training: provides visual features
- Restoration examples → Architecture search: provides search data
- Architecture search → Model selection: provides model candidates
- Model training → Model selection: provides trained variants
- Model selection → Confidence calibration: provides selected variant
- Confidence calibration → Final model lock: locks calibrated model
- Final model lock → Protected evaluation: authorizes evaluation
- Protected evaluation → Results analysis: provides test results
- Final model lock → Automatic restoration: selects deployment model
- Restoration examples → Automatic restoration: provides genuine gaps
- Automatic restoration → Restoration runtime: uses inference
- Automatic restoration → Line and tablet results: writes restorations
- Researcher → Automatic restoration: requests restoration