Pretrain stage: Synthetic data pretraining on large-scale synthetic reaction diagrams.
SFT stage: Supervised fine-tuning with three specialized tasks:
- Vanilla Reaction Parsing (VRP): Standard reaction diagram parsing
- Region-Guided Reaction Parsing (RGRP): Region-aware parsing with spatial guidance
- Prefix-Perturbed Reaction Parsing (PPRP): Robustness-oriented parsing with prefix perturbations
DPO stage: Direct Preference Optimization to further align model outputs with ground-truth annotations.

Training Details

Config	Pretrain	SFT	DPO
Base Model	Qwen2.5-VL-3B-Instruct	Qwen2.5-VL-3B-Instruct	Qwen2.5-VL-3B-Instruct
Learning Rate	1.0×10⁻⁶	1.0×10⁻⁵	3.0×10⁻⁷
Batch Size	16	4	64
Epochs	1	1	1
Scheduler	Cosine (warmup 0.03)	Cosine (warmup 0.03)	Cosine (warmup 0.03)
Optimizer	AdamW	AdamW	AdamW
Trainable Params	LLM only	Full (vision + LLM)	LLM only

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Jingcz/RobustRDP"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

Related Resources

Dataset: RxnLabelData
Annotation Platform: RxnLabel