Upload README.md with huggingface_hub

README.md

@@ -1,6 +1,6 @@
 ---
 viewer: false
-tags: [uv-script, ocr, vision-language-model, document-processing]
+tags: [uv-script, ocr, vision-language-model, document-processing, hf-jobs]
 ---
 
 # OCR UV Scripts
@@ -70,6 +70,50 @@ hf jobs uv run --flavor l4x1 \
     --batch-size 32
 ```
 
+### GLM-OCR (`glm-ocr.py`) 🏆 SOTA on OmniDocBench V1.5!
+
+Compact high-performance OCR using [zai-org/GLM-OCR](https://huggingface.co/zai-org/GLM-OCR) with 0.9B parameters:
+
+- 🏆 **94.62% on OmniDocBench V1.5** - #1 overall ranking
+- 🧠 **Multi-Token Prediction** - MTP loss + stable full-task RL for quality
+- 📝 **Text recognition** - Clean markdown output
+- 📐 **Formula recognition** - LaTeX mathematical notation
+- 📊 **Table recognition** - Structured table extraction
+- 🌍 **Multilingual** - zh, en, fr, es, ru, de, ja, ko
+- ⚡ **Compact** - Only 0.9B parameters, MIT licensed
+- 🔧 **CogViT + GLM** - Visual encoder with efficient token downsampling
+
+**Task Modes:**
+
+- `ocr`: Text recognition (default)
+- `formula`: LaTeX formula recognition
+- `table`: Table extraction
+
+**Quick start:**
+
+```bash
+# Basic OCR
+hf jobs uv run --flavor a100-large \
+    -s HF_TOKEN \
+    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/glm-ocr.py \
+    your-input-dataset your-output-dataset \
+    --max-samples 100
+
+# Formula recognition
+hf jobs uv run --flavor a100-large \
+    -s HF_TOKEN \
+    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/glm-ocr.py \
+    scientific-papers formulas-extracted \
+    --task formula
+
+# Table extraction
+hf jobs uv run --flavor a100-large \
+    -s HF_TOKEN \
+    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/glm-ocr.py \
+    documents tables-extracted \
+    --task table
+```
+
 ### LightOnOCR (`lighton-ocr.py`) ⚡ Good one to test first since it's small and fast!
 
 Fast and compact OCR using [lightonai/LightOnOCR-1B-1025](https://huggingface.co/lightonai/LightOnOCR-1B-1025):
@@ -319,6 +363,14 @@ hf jobs uv run --flavor l4x1 \
     --task-mode formula \
     --batch-size 32
 
+# GLM-OCR - SOTA 0.9B model (94.62% OmniDocBench)
+hf jobs uv run --flavor a100-large \
+    -s HF_TOKEN \
+    https://huggingface.co/datasets/uv-scripts/ocr/raw/main/glm-ocr.py \
+    your-input-dataset your-output-dataset \
+    --batch-size 16 \
+    --max-samples 100
+
 # DeepSeek-OCR - Real-world example (National Library of Scotland handbooks)
 hf jobs uv run --flavor a100-large \
     -s HF_TOKEN \
@@ -482,6 +534,13 @@ Any HuggingFace dataset containing images - documents, forms, receipts, books, h
 - `--output-column`: Override default column name (default: `paddleocr_[task_mode]`)
 - Ultra-compact 0.9B model - fastest initialization!
 
+**GLM-OCR**:
+
+- `--task`: Task type - `ocr` (default), `formula`, or `table`
+- `--repetition-penalty`: Repetition penalty (default: 1.1, from official SDK)
+- Near-greedy sampling by default (temperature=0.01, top_p=0.00001) matching official SDK
+- Requires vLLM nightly + transformers from git (handled automatically)
+
 **DeepSeek-OCR**:
 
 - `--resolution-mode`: Quality level - `tiny`, `small`, `base`, `large`, or `gundam` (default)