GravityOCR-KO (internal, private)

Korean continued fine-tuning of trillionlabs/GravityOCR, with replay from the original English/CJK training pools so the table and formula tasks are retained.

Like the base model, this is a standard GlmOcrForConditionalGeneration checkpoint plus block_diffusion.json. ar_loss_weight: 1.0 is preserved, so the checkpoint still has a trained AR path and can self-speculate (diffusion drafts, AR verifies).

Training

base	trillionlabs/GravityOCR (post-GRPO release, not the GLM-OCR base)
objective	joint AR + block diffusion, ar_loss_weight=1.0, bd_size=32, block-phase jitter on
caps	max_length=16384, max_response_length=6144, max_vision_tokens=8192

Usage

import torch
from transformers import AutoProcessor, GlmOcrForConditionalGeneration

repo = "trillionlabs/GravityOCR-KO"
model = GlmOcrForConditionalGeneration.from_pretrained(repo, dtype=torch.bfloat16, device_map="cuda")
processor = AutoProcessor.from_pretrained(repo)

Stock transformers>=5.8 loads it for AR decoding. Self-speculative decoding needs the nanoVLM-DiffusionOCR repository (in process) or the SGLang patch (serving).