MolSeek-OCR

ChemSeek-OCR

Downloading Source Code

mkdir llm4chem
cd llm4chem
git clone https://github.com/deepseek-ai/DeepSeek-OCR-2.git
git clone https://github.com/HaCTang/ChemSeek-OCR.git

Environment Setting

First build a conda environment following https://github.com/deepseek-ai/DeepSeek-OCR-2.

conda create -n chemseek-ocr python=3.12.9 -y
conda activate chemseek-ocr

Download the vllm-0.8.5 whl. The file is at the bottom of the page.

cd DeepSeek-OCR-2
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu118
pip install vllm-0.8.5+cu118-cp38-abi3-manylinux1_x86_64.whl
pip install -r requirements.txt
pip install flash-attn==2.7.3 --no-build-isolation
pip install peft accelerate wandb
pip install matplotlib albumentations opencv-python rdkit SmilesPE pandas

Note: if you want vLLM and transformers codes to run in the same environment, you don't need to worry about this installation error like: vllm 0.8.5+cu118 requires transformers>=4.51.1

Download Weight and Test DeepSeek-OCR-2

Assume that you are still in the directory of DeepSeek-OCR-2

export HF_HOME=$PWD/hf_cache
export TRANSFORMERS_CACHE=$PWD/hf_cache
export HF_DATASETS_CACHE=$PWD/hf_cache
export HF_HUB_DISABLE_XET=1

You can do transformer-based inference and download the weight of DeepSeek-OCR-2 at the same time:

from transformers import AutoModel, AutoTokenizer
import torch
import os
os.environ["CUDA_VISIBLE_DEVICES"] = '1'
model_name = 'deepseek-ai/DeepSeek-OCR-2'
# model_name = 'model_file'

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModel.from_pretrained(model_name, _attn_implementation='flash_attention_2', use_safetensors=True, trust_remote_code=True)
model = model.eval().cuda().to(torch.bfloat16)

# prompt = "<image>\nFree OCR. "
# prompt = "<image>\n<|grounding|>Give me the smiles of the molecule. "
prompt = "<image>\n Give me the smiles of the molecule. "
image_file = 'test_img/penicillin.jpg'
output_path = 'output'

res = model.infer(tokenizer, prompt=prompt, image_file=image_file, output_path = output_path, base_size = 1024, image_size = 768, crop_mode=True, save_results = True)

Training Datasets

cd ../ChemSeek-OCR

Download the following benchmarks and unzip them into the folder training_data. pubchem uspto-mol

You can choose to use dataset.py to pre-render pictures in pubchem.

Evaluation Checkpoint

The released weights corresponding to the reported results are hosted on Hugging Face Hub under Haocheng1/MolSeek-OCR:

• LoRA weight: the LoRA SFT checkpoint after merging into the base DeepSeek-OCR-2 model.
• Progressive Full SFT weight: the progressive full-parameter SFT checkpoint initialized from the merged LoRA model.

These fine-tuned weights follow the license and availability conditions of the base DeepSeek-OCR-2 model.

Download Benchmarks

Download the following benchmarks and unzip them into the folder benchmark. Synthetic Realistic Perturbed

Run Evaluation

Please use chemseek-ocr for evaluation! Don't use chemseek-ocr-verl

Check evaluation.py and evaluation_config.yaml.

python evaluation.py --config evaluation_config.yaml
python gspo/evaluation_gspo.py --config gspo/evaluation_gspo_config.yaml

Progressive SFT

The training of Progressive SFT consists of three parts: cold-start LORR SFT in a few-sample scenario, then merging the LoRA weights into the base model, and finally performing full-parameter SFT.

LoRA SFT

See File lora_sft.py and lora_sft_config.yaml.

Default setting using 3*64k datapoints to do cold-start LoRA.

For training dataset setting in lora_sft_config.yaml, uspto-mol uses realistic data mode, while pubchem uses dynamic data mode. And pubchem can actually use augmentation. For example:

train_sets:
 - train_csv: ./training_data/pubchem/train_200k.csv
   data_mode: dynamic
   pre_rendered_image_dir: null
   realistic_image_root: null
   instruction: "<image>\n Give me the SMILES of the molecule. "
   style: molscribe_default # molscribe_default / chemdraw_like
   mol_augment: true
   include_condensed: False
   max_samples: null
   sample_num: 64000
 - train_csv: ./training_data/pubchem/train_200k.csv
   data_mode: dynamic
   pre_rendered_image_dir: null
   realistic_image_root: null
   instruction: "<image>\n Give me the SMILES of the molecule. "
   style: chemdraw_like # molscribe_default / chemdraw_like
   mol_augment: false
   include_condensed: False
   max_samples: null
   sample_num: 64000
 - train_csv: ./training_data/uspto_mol/train_200k.csv
   data_mode: realistic
   realistic_image_root: ./training_data
   instruction: "<image>\n Give me the SMILES of the molecule. "
   sample_num: 64000

For other parameters, you can change:

...
batch_size: 4 -> 32
grad_accum: 8 -> 1
...
accelerate_num_processes: 4 -> 8
accelerate_gpu_ids: "0,1,2,3" -> "0,1,2,3,4,5,6,7,8"
...

You can run lora_sft.py with:

python lora_sft.py --config lora_sft_config.yaml

Merging LoRA weight

python merge_lora_weight.py \
  --pretrained_weight_path ../DeepSeek-OCR-2 \
  --checkpoint_path ./weight/checkpoint-1500 \
  --merged_model_dir ./merged_models \
  --full_or_lora lora

Full parameter SFT

Similar to LoRA SFT session, for training dataset setting in lora_sft_config.yaml, uspto-mol uses realistic data mode, while pubchem uses dynamic data mode. And pubchem can actually use augmentation. For example:

train_sets:
 - train_csv: ./training_data/pubchem/train_1m.csv
   data_mode: dynamic
   pre_rendered_image_dir: null
   realistic_image_root: null
   instruction: "<image>\n Give me the SMILES of the molecule. "
   style: molscribe_default # molscribe_default / chemdraw_like
   mol_augment: true
   include_condensed: False
   max_samples: null
   sample_num: null
 - train_csv: ./training_data/pubchem/train_1m.csv
   data_mode: dynamic
   pre_rendered_image_dir: null
   realistic_image_root: null
   instruction: "<image>\n Give me the SMILES of the molecule. "
   style: chemdraw_like # molscribe_default / chemdraw_like
   mol_augment: false
   include_condensed: False
   max_samples: null
   sample_num: null
 - train_csv: ./training_data/uspto_mol/train_680k.csv
   data_mode: realistic
   realistic_image_root: ./training_data
   instruction: "<image>\n Give me the SMILES of the molecule. "
   sample_num: null

The total batch size can be a little bit larger, around 512-2048. Learning rate is around 1/10 to 1/100 compared with previous LoRA, like 5e-6. Run 1-2 epochs (you should evaluate the steps).

You can run progressive_sft.py with:

python progressive_sft.py --config progressive_sft_config.yaml

RL

The RL phase is divided into three progressive stages to ensure the stability of the MoE architecture while optimizing for chemical SMILES accuracy.

Download the verl-0.6.1 Source Code. The file is at the bottom of the page.

conda create -n chemseek-ocr-verl python=3.12
conda activate chemseek-ocr-verl
unzip verl-0.6.1.zip
cd verl-0.6.1

In order to be compatible with the fine-tuning code of DeepSeek-OCR-2, please open verl-0.6.1/setup.py. Set "transformers==4.57" in the install_requires, and set ["tensordict>=0.8.0,<=0.10.0,!=0.9.0", "vllm==0.8.5"] in the VLLM_REQUIRES.

# pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu118 #You can run this if error says libcudart.so.11.0
pip install -e .[vllm]
pip install flash-attn==2.7.3 --no-build-isolation
pip install matplotlib albumentations rdkit SmilesPE pandas addict

GSPO RL

GSPO (Group Sequence Policy Optimization) uses tight symmetric clipping (clip_ratio_low/clip_ratio_high) instead of KL regularization, providing more stable policy updates. The implementation is built on the verl framework and is now organized under the gspo/ directory:

Key GSPO parameters in the config (gspo: section):

• clip_ratio_low / clip_ratio_high: Tight symmetric clipping bounds (default 3e-4 / 4e-4) that replace KL penalty.
• clip_ratio_c: Upper clip bound for the importance-sampling ratio (default 10.0).
• loss_agg_mode: seq-mean-token-mean: Sequence-level then token-level mean aggregation.
• use_kl_loss: false / kl_loss_coef: 0.0: KL loss is disabled; tight clipping suffices.
• group_size: 8: Number of responses sampled per prompt for advantage estimation.
• use_dynamic_bsz: true: Dynamic batch sizing based on token budget.

The reward function computes five SMILES-based components (validity, tanimoto, canon_smiles, graph, chiral) with configurable weights, normalized to [0, 1] by dividing by the weight sum.

# Step 1: Prepare parquet data
python gspo/prepare_verl_data.py --config gspo/gspo_rl_verl_config.yaml --workers 8

# Step 2: Launch GSPO training
python gspo/gspo_rl_verl.py --config gspo/gspo_rl_verl_config.yaml

# Step 3: Evaluate a GSPO checkpoint
python gspo/evaluation_gspo.py --config gspo/evaluation_gspo_config.yaml

ReFT

ReFT (Rejection sampling Fine-Tuning) is organized under the reft/ directory and provides a two-phase pipeline: best-of-N generation with vLLM followed by SFT on the curated high-reward samples.

# Run the full ReFT pipeline
python reft/reft.py --config reft/reft_config.yaml

# Only generate best-of-N candidates
python reft/reft.py --config reft/reft_config.yaml --phase generate

# Only train a specific iteration
python reft/reft.py --config reft/reft_config.yaml --phase train --iteration 0

Inference

For transformer inference, please use conda environment chemseek-ocr:

python transformer_infer_case.py --model-path ./weight_progressive_sft/checkpoint-2000 --image-file ./test_img/penicillin.jpg

Citation

@misc{tang2026finetuningdeepseekocr2molecularstructure,
      title={Fine-tuning DeepSeek-OCR-2 for Molecular Structure Recognition}, 
      author={Haocheng Tang and Xingyu Dang and Junmei Wang},
      year={2026},
      eprint={2604.03476},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.03476}, 
}