omarelsherif010/glm-ocr-bnk-finetuning

GLM-OCR Fine-Tuning Pipeline

Fine-tuning GLM-OCR 0.9B (CogViT encoder + GLM-0.5B decoder) for Korean financial document table recognition using LoRA via LLaMA-Factory.

Performance Targets

Metric	Target
TEDS (2-level nested)	>= 90%
TEDS (3-level nested)	>= 85%
Korean CER	<= 1%
Latency	<= 0.5s/page

Directory Structure

glm_ocr_finetuning/
├── config/                          # Training/eval YAML configs
│   ├── training_config.yaml         # LoRA + LLaMA-Factory config
│   ├── preprocessing_config.yaml
│   └── evaluation_config.yaml
│
├── data/                            # ALL raw source data (gitignored)
│   ├── existing/                    #   216 pre-labeled BNK images+labels
│   │   ├── images/*.png
│   │   └── labels/*.html
│   ├── dart/                        #   9,375 DART financial tables
│   │   ├── images/*.png
│   │   └── labels/*.html
│   ├── fss/                         #   209 FSS tables (flat png+html)
│   ├── wiki/                        #   299 Korean Wikipedia tables (flat png+html)
│   ├── synthetic/                   #   1,000 Jinja2-generated tables
│   │   ├── images/*.png
│   │   └── labels/*.html
│   └── synthetic-shilu/             #   3,000 BNK financial tables + metadata.json
│
├── data_collection/                 # Data collection & formatting scripts
│   ├── format_training_data.py      # Unified formatter (all sources -> JSONL)
│   ├── format_existing_dataset.py   # Legacy formatter for existing dataset
│   ├── scrape_dart.py               # DART financial reports scraper
│   ├── scrape_fss_stats.py          # FSS statistics scraper
│   ├── scrape_wiki_tables.py        # Korean Wikipedia table scraper
│   └── synthetic/                   # Synthetic data generation
│       ├── generate_synthetic.py    # Playwright-based table renderer
│       ├── data_pools.py            # Korean vocabulary pools
│       └── templates/               # 16 Jinja2 HTML templates
│
├── dataset/                         # Pipeline output (train/val/test splits)
│   ├── train.jsonl                  # 12,017 training entries
│   ├── val.jsonl                    # 2,123 validation entries
│   ├── test.jsonl                   # 33 test entries
│   ├── training_data.jsonl          # Combined 14,140 entries
│   ├── dataset_stats.csv            # Per-entry statistics
│   └── llamafactory/               # LLaMA-Factory ShareGPT format (gitignored)
│       ├── train_sharegpt.json
│       ├── val_sharegpt.json
│       └── dataset_info.json
│
├── preprocessing/
│   ├── preprocess.py                # Image preprocessing (DPI, deskew, CLAHE)
│   └── postprocess.py               # HTML postprocessing (repair, Korean correction)
│
├── evaluation/
│   ├── metrics/
│   │   ├── teds.py                  # Tree-Edit-Distance-based Similarity
│   │   ├── cer_wer.py               # Character/Word Error Rate
│   │   ├── table_f1.py              # Table F1 score
│   │   └── nesting_depth.py         # Nesting depth accuracy
│   └── test_set/                    # Curated test images (TBD)
│
├── training/
│   ├── train.py                     # Main training entry point
│   ├── direct_trainer.py            # Direct HuggingFace trainer
│   └── convert_dataset.py           # JSONL -> LLaMA-Factory converter
│
├── evaluate.py                      # Evaluation runner
└── outputs/                         # Training outputs (gitignored)

Dataset Summary

14,099 unique entries (14,140 after oversampling for balance), from 6 sources:

Source	Count	Description
existing	216	Pre-labeled BNK reference dataset (augmented)
dart	9,375	DART financial disclosure tables (scraped)
fss	209	FSS statistical tables
wiki	299	Korean Wikipedia tables
synthetic	1,000	Jinja2-generated tables (Playwright rendered)
synthetic-shilu	3,000	BNK financial tables (16 templates, 2/3-level + merged)

Category distribution:

Category	Count	%
plain	3,675	26.0%
merged (rowspan/colspan)	9,300	65.8%
nested_2level	715	5.1%
nested_3level	450	3.2%

Splits: Train 12,017 / Val 2,123 / Test 33

Quick Start

1. Format training data (all sources)

backend/.venv/bin/python backend/scripts/glm_ocr_finetuning/data_collection/format_training_data.py \
  --sources all \
  --output backend/scripts/glm_ocr_finetuning/dataset/

2. Convert to LLaMA-Factory format

backend/.venv/bin/python backend/scripts/glm_ocr_finetuning/training/convert_dataset.py \
  --input backend/scripts/glm_ocr_finetuning/dataset/train.jsonl \
         backend/scripts/glm_ocr_finetuning/dataset/val.jsonl \
  --output backend/scripts/glm_ocr_finetuning/dataset/llamafactory/

3. Dry-run training

backend/.venv/bin/python backend/scripts/glm_ocr_finetuning/training/train.py --dry-run

4. Train on GPU (requires LLaMA-Factory + GPU)

pip install llamafactory
backend/.venv/bin/python backend/scripts/glm_ocr_finetuning/training/train.py --mode llamafactory

JSONL Schema

Each entry in train.jsonl / val.jsonl:

{
  "image_path": "backend/scripts/glm_ocr_finetuning/data/dart/images/dart_xxx.png",
  "ground_truth_html": "<table><tr><td>...</td></tr></table>",
  "nesting_level": 1,
  "source": "dart",
  "category": "merged",
  "metadata": {
    "original_filename": "dart_xxx.png",
    "validation": {
      "table_count": 1,
      "row_count": 15,
      "cell_count": 60,
      "has_rowspan": true,
      "has_colspan": true,
      "text_length": 500
    }
  }
}

Training Configuration

From config/training_config.yaml:

Parameter	Value
Base model	zai-org/GLM-OCR
LoRA rank	8
LoRA alpha	16
LoRA targets	all linear layers
Batch size	2 (effective 16 with grad accum)
Learning rate	1e-4 (cosine schedule)
Epochs	3
Precision	bf16
Max seq length	4096

Adding New Data Sources

1. Place raw data under data/<source_name>/ (flat *.png + *.html pairs, or images/ + labels/ subdirs)
2. Add a loader in data_collection/format_training_data.py:
   • Flat structure (png+html side by side): use load_flat_source()
   • Subdirectory structure (images/ + labels/): use load_scraped_source()
3. Register in SOURCE_LOADERS dict
4. Add to argparse --sources choices
5. Re-run the pipeline (steps 1-3 from Quick Start)

Data Collection Scripts

Script	Source	Output	Notes
scrape_dart.py	DART API	data/dart/	Uses OpenDART API, renders via Playwright
scrape_fss_stats.py	FSS Korea	data/fss/	Flat png+html pairs
scrape_wiki_tables.py	Korean Wikipedia	data/wiki/	Flat png+html pairs
synthetic/generate_synthetic.py	Jinja2 templates	data/synthetic/	16 templates, async Playwright
format_training_data.py	All sources	dataset/	Unified formatter with validation + dedup

Evaluation

# Run evaluation on test set with live model
backend/.venv/bin/python backend/scripts/glm_ocr_finetuning/evaluate.py \
  --test-set backend/scripts/glm_ocr_finetuning/dataset/test.jsonl \
  --live --endpoint http://localhost:8001

Key Design Decisions

1. Single data/ directory: All 6 raw data sources under one gitignored directory for consistency
2. GLM-OCR 0.9B: Lightweight vision-language model purpose-built for table recognition
3. LLaMA-Factory: Standardized SFT framework with ShareGPT format for easy training
4. LoRA (rank 8): Efficient fine-tuning, fits on consumer GPUs
5. Korean-only training data: Optimized for Korean financial documents
6. Flat + subdirectory loaders: Flexible data ingestion supporting both file layouts

Team

• Omar: Model infrastructure, training pipeline, preprocessing/postprocessing
• Shilu: Data collection (FSS, Wikipedia, synthetic-shilu 3,000 tables)

References

• GLM-OCR - Base model
• LLaMA-Factory - Training framework
• TEDS - Primary evaluation metric
• LoRA - Parameter-efficient fine-tuning