This is the intermediate SFT checkpoint before GRPO refinement. The best model with GRPO is available at blazeofchi/pdf-ocr-rl-qwen3vl2b-sft-grpo.

This is a LoRA adapter (r=32, alpha=64, 2.18% trainable parameters). Load it on top of the base model using PEFT.

Training Details

This checkpoint was produced by Stage 1 of the two-stage training pipeline.

SFT Training (100 steps)

Teaches the model the image-to-markdown mapping using supervised examples from rendered PDF pages paired with their source markdown.

Parameter	Value
Learning rate	2e-5
Batch size	2
Max steps	100
Framework	Unsloth + TRL SFTTrainer
Loss curve	1.295 → 0.78
Grad norm	~1.85

Technical Details

Detail	Value
Base model	unsloth/Qwen3-VL-2B-Instruct (2.15B params)
LoRA rank (r)	32
LoRA alpha	64
LoRA dropout	0.0
Target modules	All linear layers
Trainable parameters	2.18% of total
Precision	bf16
Hardware	NVIDIA A40 48GB (RunPod)
Dataset	500 train samples from blazeofchi/pdf-ocr-rl-dataset

Purpose

This checkpoint serves two purposes:

Intermediate checkpoint for GRPO: This is the starting point for Stage 2 (GRPO refinement). GRPO alone produces near-zero gradients on vision-language models — SFT warm-up is essential.
SFT-only baseline: Compare against the full SFT+GRPO model to measure the contribution of GRPO refinement.

Usage

python

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from peft import PeftModel
from qwen_vl_utils import process_vision_info
from PIL import Image
import torch

base_model = Qwen3VLForConditionalGeneration.from_pretrained(
    "unsloth/Qwen3-VL-2B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

model = PeftModel.from_pretrained(
    base_model, "blazeofchi/pdf-ocr-rl-qwen3vl2b-sft-only"
)
processor = AutoProcessor.from_pretrained(
    "blazeofchi/pdf-ocr-rl-qwen3vl2b-sft-only"
)

image = Image.open("page.png")
messages = [
    {"role": "user", "content": [
        {"type": "image", "image": image},
        {"type": "text", "text": "Convert this PDF page to well-structured markdown."}
    ]}
]

text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text], images=image_inputs,
    padding=True, return_tensors="pt"
).to(model.device)

output = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
result = processor.decode(
    output[0][inputs.input_ids.shape[1]:], skip_special_tokens=True
)
print(result)

Related Models

Model	Description
blazeofchi/pdf-ocr-rl-qwen3vl2b-sft-grpo	SFT + GRPO (best model)
blazeofchi/pdf-ocr-rl-qwen3vl2b-sft-only	SFT-only (this model)

Citation

bibtex

@misc{pdf-ocr-rl-2026,
  title={PDF-OCR-RL: Fine-tuning Vision-Language Models for PDF-to-Markdown with GRPO},
  author={Paras Sharma},
  year={2026},
  url={https://github.com/Parassharmaa/pdf-ocr-rl}
}

License

Apache 2.0 (same as base model)