Hugging Face's logo

Datasets:
Jingcz
/
RobustRDP-ProcessedTrainData

Tasks:
Image-to-Text
Visual Question Answering
Sub-tasks:
visual-question-answering
Languages:
English
Size:
100K<n<1M
License:
Dataset card Files
xet
 Community
RobustRDP-ProcessedTrainData / README.md
Jingcz's picture
Jingcz
Update README.md
091ca79 verified
preview code
|
Raw
History Blame Contribute Delete
3.81 kB
metadata 
annotations_creators:
  - machine-generated
language:
  - en
license:
  - other
multilinguality:
  - monolingual
size_categories:
  - 100K<n<1M
task_categories:
  - image-to-text
  - visual-question-answering
task_ids:
  - visual-question-answering

RobustRDP-ProcessedTrainData

Processed training data for the paper: RobustRDP: Advancing Reaction Diagram Parsing via Synthetic-to-Real Data Scaling and Robustness-Oriented Training.

Dataset Structure 

pretrain_data/
├── pretrain_downsampled_llm_6w.json   # 60,000 synthetic reaction diagrams
└── images_pretrain_resized/
    ├── single_line_resized/           # Single-line chain-style reactions
    ├── multi_line_resized/            # Multi-line chain-style reactions
    ├── branch_resized/                # Branching reactions
    └── cycle_resized/                 # Cyclic reactions

sft_data/
├── multi_task_sft_downsampled_llm.json # 190,800 multi-task SFT entries
├── images_train_aug_resized/          # Augmented training images
└── images_train_resized/              # Original resized training images

dpo_data/
└── train_downsampled_llm_dpo.json     # 14,169 DPO triplets (chosen/rejected pairs)

Data Splits

Split Entries Description
Pretrain 60,000 Synthetic reaction diagrams generated by LayoutDrivenSynthesizer (4 layout types: single-line, multi-line, branch, cycle)
SFT 190,800 Multi-task supervised fine-tuning data with 3 task variants: Vanilla Reaction Parsing (VRP), Region-Guided Reaction Parsing (RGRP), Prefix-Perturbed Reaction Parsing (PPRP)
DPO 14,169 Direct Preference Optimization triplets with chosen (ground-truth) and rejected (model prediction) annotations

Data Format

Pretrain & SFT

Each entry follows the conversational format:

{
    "messages": [
        {"content": "<image>\n...", "role": "user"},
        {"content": "<rxn><rct>...<mol><cnd>...<txt><prd>...<mol>", "role": "assistant"}
    ],
    "images": ["path/to/image.png"]
}

Annotations use special tokens:

  • <rxn>: Start of a reaction
  • <rct>: Reactants section
  • <cnd>: Conditions section
  • <prd>: Products section
  • <mol>: Molecule entity
  • <txt>: Text entity

DPO

Each entry contains chosen/rejected pairs:

{
    "messages": [{"from": "user", "value": "<image>\n..."}],
    "chosen": {"from": "assistant", "value": "<rxn>..."},
    "rejected": {"from": "assistant", "value": "<rxn>..."},
    "images": ["path/to/image.png"],
    "overall": {"precision": 0.8, "recall": 0.8, "f1": 0.8},
    "mol_only": {"precision": 0.9, "recall": 0.9, "f1": 0.9}
}

Data Generation

  • Pretrain: Synthetic data generated by LayoutDrivenSynthesizer — renders molecules from PubChem SMILES onto reaction diagrams with 4 layout types.
  • SFT: Built from RxnLabelData with data augmentation (rotation, distortion, composite) and two auxiliary tasks (RGRP, PPRP). See SFT data process.
  • DPO: Generated by running the SFT model on training data and filtering cases where model predictions (F1 < 0.8) differ from ground truth. See DPO data process.

Related Resources