phoenix-arabic-manuscript-htr / METRICS_TABLE.md
factlogic's picture
Update publication links for Phoenix model name
af3342d verified
|
Raw
History Blame
4.11 kB

📊 VERIFIED METRICS TABLE: Rakīm AI (رَقِيم)

Below is the verified dataset of digital metrics compiled from the project's verification sample reports, test logs, presentation JSON configurations, and service scripts.


1. Character Error Rate (CER) and Model Metrics

Metric Value Dataset / Sample Source File Confidence Notes
Trained Sample CER 3.29% (96.71% Acc) 3 training pages (Maghrebi) عينات_التحقق/النتائج.md High Measures how well the model memorizes its training subset.
Unseen Test CER 7.50% (92.50% Acc) 51 frozen test pages (Maghrebi) عينات_التحقق/النتائج.md High Primary metric for general readability and model performance.
Baseline exp6 (No LM) CER 7.48% RASAM Maghrebi dataset presentation_dashboard/data/metrics.json High Verified in direct comparison logs.
LM exp6 (With LM) CER 6.79% RASAM Maghrebi dataset presentation_dashboard/data/metrics.json High Shows a 0.69% absolute error reduction using local LM.
Baseline TariMa (No LM) CER 10.27% TariMa manuscript dataset presentation_dashboard/data/metrics.json High Generalization baseline for TariMa hands.
LM TariMa (With LM) CER 8.81% TariMa manuscript dataset presentation_dashboard/data/metrics.json High Shows a 1.46% absolute error reduction using local LM.
Legacy Model CER Unusable (Garbage Text) Maghrebi scripts المقارنة_الموحدة.md High Model reg(96.5) fails completely on historical Maghrebi hands.

2. Segmentation and Page Layout Metrics

Page/Sample Ground Truth (GT) Lines Produced Lines (seg_best) Source File Status / Analysis
Page 1926 (Simple Layout) 9 lines 11 lines عينات_التحقق/النتائج.md Close to perfect segment alignment.
Page 417 (Margins/Hawahin) 13 lines 43 lines عينات_التحقق/النتائج.md Severe over-segmentation due to marginalia.
Page 1982 (Dense + Margins) 37 lines 53 lines عينات_التحقق/النتائج.md Over-segmentation of page borders/marginal notes.
Page 1960 (Layout Split) 26 lines (Main) 30 lines (26 Main / 4 Margins) demo_package/المقارنات_الرقمية.md Layout split correctly separates margins from text body.

3. Database Catalog Statistics

Metric Value Source File Status
Total Cataloged Works 35,632 catalog_stats.json Compiled from OpenITI (13,364) & FIHRIST (15,543).
Total Manuscript Copies 49,980 catalog_stats.json Total registered manuscript copies.
Works with Multiple Copies 6,003 (16.85%) catalog_stats.json Works residing in more than one library collection.
Average Copies per Work 1.4 catalog_stats.json Global duplication factor.

4. Execution Latency Metrics (CPU vs. GPU)

Task Device Time (Seconds) Source File / Logs Notes
Neural Segmentation (BLLA) CPU (Docker) ~11.0s demo_package/README_DEMO.md BLLA is structurally heavy.
Text Recognition CPU (Docker) ~4.0s demo_package/README_DEMO.md Direct Kraken PyTorch execution.
Total Inference Page Latency CPU (Docker) ~14.0–16.0s demo_package/README_DEMO.md Time required to process a single raw page.
Fast Segmentation CPU (Direct) 1.0–3.0s config.py Settings flag FAST_SEGMENTATION=true.
Total Inference Page Latency GPU (RTX 4050) ~4.0s demo_package/README_DEMO.md Substantial acceleration (almost 4x speedup).
Inference Page Latency Cache Hit 0.0s ocr.py Immediate response if image hash matches cache database.