tieubaoca/glm-ocr-viet-mixed-lora

GLM-OCR Vietnamese LoRA Adapter

LoRA adapter for GLM-OCR fine-tuned on mixed Vietnamese OCR data.

Training Data

• 5CD-AI/Viet-Handwriting-OCR-v2
• ademax/ocr_scan_vi_01

Runtime

Use with llama.cpp / llama-server as:

llama-server \
  -hf ggml-org/GLM-OCR-GGUF \
  --lora adapter-f16.gguf

Artifacts

• adapter-f16.gguf: GGUF LoRA adapter used by inference
• adapter_model.safetensors: original HF LoRA weights
• adapter_config.json: PEFT adapter config

Training Metrics

• train_runtime: 10392.5138
• train_loss: 0.6893482752761401
• train_samples_per_second: 20.279
• train_steps_per_second: 1.268
• epoch: 3.0

Eval Metrics

• eval_loss: 0.3983404338359833
• eval_runtime: 69.8699
• eval_samples_per_second: 28.625
• eval_steps_per_second: 28.625
• epoch: 3.0

All Results

• epoch: 3.0
• eval_loss: 0.3983404338359833
• eval_runtime: 69.8699
• eval_samples_per_second: 28.625
• eval_steps_per_second: 28.625
• total_flos: 2.4598838533767168e+17
• train_loss: 0.6893482752761401
• train_runtime: 10392.5138
• train_samples_per_second: 20.279
• train_steps_per_second: 1.268

Notes

• Adapter is published separately and loaded at runtime from the Hugging Face repo.
• Base model remains ggml-org/GLM-OCR-GGUF.