📜 Chronos-VL: The 1545 Resurrection Engine

“We don't just read history. We translate time.”

Chronos-VL is a specialized Vision–Language Model system designed to decipher Early Modern Spanish Gothic manuscripts (c. 1545).
Built for the Baidu ERNIE AI Developer Challenge.

---

📉 The Problem

Millions of pages of legal, social, and cultural history remain locked in Spanish archives (e.g., The RODRIGO Corpus).

• Standard OCR (Tesseract / Base Models)
  Fails catastrophically (≈68% error rate) due to:

  • Ink bleed-through
  • Dense ligatures
  • Gothic calligraphy
  • The infamous Long S (ſ / f) ambiguity

• The Gap
  Historians need searchable, modernized text, not just noisy transcriptions.

---

💡 The Solution: Chronos-VL System

Chronos-VL introduces a two-stage vision–language pipeline purpose-built for 16th-century manuscripts.

1️⃣ Visual Perception (AI Layer)

• Fine-tuned PaddleOCR-VL-0.9B
• Training via ERNIEKit (SFT)
• Optimized on NVIDIA A100 (80GB)
• Learns period-specific Gothic features and ligatures

2️⃣ Semantic Modernization (Logic Layer)

• Chronos Engine post-processes OCR output
• Normalizes archaic spelling:
  • dixo → dijo
  • facer → hacer
• Produces:
  • Clean modern Spanish

---

📊 Performance Benchmarks

A/B testing conducted on 100 unseen manuscript pages.

Metric	PaddleOCR (Base)	Chronos-VL (Ours)	Improvement
Median Character Error Rate	19.82%	1.64%	12× Better
Usable Output (<5% Error)	1%	77%	76× Increase
Word Error Rate	74.44%	17.35%	4× Better

---

🚀 Interactive Demo

Upload any 16th-century manuscript and see Chronos-VL in action.

---

📂 Repository Structure

Here is a breakdown of the core files in this repository:

File	Description
📓 Demo_Chronos_VL.ipynb	The Interactive App. A complete Colab notebook that launches the Gradio interface, allowing you to upload images and compare baseline and finetuned model.
📓 Evaluation.ipynb	The Proof. The script used to benchmark the model against 100 unseen images. Generates the CER/WER statistics and comparisons.
🐍 chronos_processing.py	The Logic Layer. Contains the custom ChronosPostProcessor class for hallucination filtering, archaic text modernization.
📄 Finetuning_script.txt	The Training Protocol. The exact commands and configurations used with ERNIEKit to train the model on the NVIDIA A100 GPU.
🖼️ Rodrigo_*.png	Sample Data. Authentic 1545 manuscript fragments from the test set. You can use these to test the demo immediately.
📊 training_loss.jpg	Convergence Metrics. Visual evidence of the training process showing the reduction in loss over 400 steps.
🖼️ finetune_success.jpg	Visual Evidence. Showing how well finetuned model is performing.

---

🧠 Technical Architecture

• Base Model: PaddleOCR-VL-0.9B (Vision-Language)
• Training Framework: ERNIEKit (SFT)
• Hardware: NVIDIA A100 (80GB)
• Dataset: RODRIGO Corpus (1545) - 9,000 Text Lines

---

🎥 Project Video

[ https://www.youtube.com/watch?v=PaK24VT_3Jk ]

🙏 Acknowledgements

• Baidu PaddlePaddle Team for the ERNIEKit framework.
• Universitat Politècnica de València for the RODRIGO dataset.