GLM_OCR-GGUF

ctogaurav/GLM_OCR-GGUF

GLM_OCR-GGUF is a machine learning model from ctogaurav. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as mit.

GGUF conversions of ctogaurav/GLMOCR — fine-tuned LoRA adapters merged into zai-org/GLM-OCR (0.9B) and quantized for fast, local, private CPU/GPU inference.

Downloads · 30 days

367

40% of all-time downloads

All-time downloads

917

Public

Repo size

3.6 GB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.gguf3.6 GB · 100%

At a glance

License: mit
Access: Public
Created: Aug 3, 2026
Updated: Sep 15, 2026
SHA: e275b4aa

Base models

zai-org/GLM-OCR

Model card

From the Hugging Face model README

GLM-OCR GGUF Models (for LM Studio / Ollama / llama.cpp)

GGUF conversions of ctogaurav/GLM_OCR — fine-tuned LoRA adapters merged into zai-org/GLM-OCR (0.9B) and quantized for fast, local, private CPU/GPU inference.

📦 Available Versions & Files

Each model requires two files — the quantized language model and its vision projector (mmproj). Download both files into the same directory:

Version	Language Model GGUF	Multimodal Projector (`mmproj`)	Ollama Setup	Status
v5.0 (Latest SOTA)	`v5.0/GLM-OCR-v5.0-Q8_0.gguf` (~682 MB)	`v5.0/mmproj-GLM-OCR-v5.0-Q8_0.gguf` (~484 MB)	`v5.0/Modelfile`	Lowest CER (0.3377), 82.0% compile
v4.1	`v4.1/GLM-OCR-v4.1-Q8_0.gguf` (~683 MB)	`v4.1/mmproj-GLM-OCR-v4.1-Q8_0.gguf` (~485 MB)	`v4.1/Modelfile`	CER: 0.3816, 82.4% compile
v3.1	`v3.1/GLM-OCR-v3.1-Q8_0.gguf` (~683 MB)	`v3.1/mmproj-GLM-OCR-v3.1-Q8_0.gguf` (~485 MB)	`v3.1/Modelfile`	Highest compile rate (88.9%)

⚡ Quickstart: Running in LM Studio

Download GLM-OCR-v5.0-Q8_0.gguf and mmproj-GLM-OCR-v5.0-Q8_0.gguf into your LM Studio models directory:
```
~/.lmstudio/models/ctogaurav/GLM_OCR-GGUF/
```
LM Studio will automatically recognize the multimodal projector (mmproj) alongside the language model.
In the Chat tab, load GLM-OCR-v5.0-Q8_0 and drag-and-drop your math scan.

🦙 Quickstart: Running in Ollama

Download the files into a folder:

huggingface-cli download ctogaurav/GLM_OCR-GGUF --include "v5.0/*" --local-dir ./glm_v5
cd glm_v5/v5.0

Create and run the model in Ollama:

ollama create glm-ocr-v5.0 -f Modelfile
ollama run glm-ocr-v5.0 "OCR this handwritten math page. Convert ONLY the handwritten mathematical content into a complete, compilable LaTeX document. Output only LaTeX. path/to/page.png"

⚠️ Recommended Inference Settings (Crucial)

Setting	Value	Rationale
Context Length (`num_ctx`)	≥ 8192	The vision encoder consumes ~1,536 tokens. Contexts below 4,096 will fail to decode.
Temperature	`0` (Greedy)	Fine-tuned and benchmarked with deterministic greedy decoding.
Repeat Penalty	`1.0` (Disabled)	Matches benchmark conditions.
Max Tokens (`num_predict`)	`1024` - `2048`	Covers full-page dense mathematical derivations.

📊 Benchmark Scorecard (250 Held-Out Pages)

System	Mean CER ↓	Norm CER ↓	Math-F1 ↑	Compile Rate ↑	Training Hardware
Base GLM-OCR	0.5151	0.4910	0.7031	0.0%	—
GLM-OCR v3.1	0.3971	0.3753	0.8171	88.9%	RTX 3060
GLM-OCR v4.1	0.3816	0.4106	0.8272	82.4%	RTX 3060
GLM-OCR v5.0 (Ours, SOTA)	0.3377 🏆	0.3683 🏆	0.8358 🏆	82.0%	Hybrid: Local RTX 3060 (step 0–250) ➔ Cloud A100 (step 250–3945)

Full training methodology, benchmark logs, and dataset curation pipeline are available at:
👉 github.com/realgauravvyas/ocr2tex \n### 🖥️ Training Journey: Local RTX 3060 ➔ Step 250 Warm-Start Cloud Handoff

Phase 1 (Local RTX 3060 12GB): Steps 0–250 trained locally (batch size 1, grad accum 8, FP16, max_length=3584, max_image_tokens=1536). Under continuous 100% compute, the GPU reached 88°C thermal throttling at 57.05s/step (~62.5 hour ETA).
Phase 2 (Cloud A100 SXM4 80GB): Handed off from the step-250 checkpoint to Lightning AI Studio, accelerating the remaining training to 3.80s/step (15× speedup) and reaching convergence loss 0.0008 in 281 minutes.\n

License: mit
Created: Aug 3, 2026
Updated: Sep 15, 2026