RobustRDP-Pretrain-SFT
Fine-tuned model for the paper: RobustRDP: Advancing Reaction Diagram Parsing via Synthetic-to-Real Data Scaling and Robustness-Oriented Training.
Description
This model is a fine-tuned checkpoint based on Qwen2.5-VL-3B-Instruct, trained with a two-stage pipeline:
- 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
Training Details
| Config | Pretrain | SFT |
|---|---|---|
| Base Model | Qwen2.5-VL-3B-Instruct | Qwen2.5-VL-3B-Instruct |
| Learning Rate | 1.0×10⁻⁶ | 1.0×10⁻⁵ |
| Batch Size | 16 | 4 |
| Epochs | 1 | 1 |
| Scheduler | Cosine (warmup 0.03) | Cosine (warmup 0.03) |
| Optimizer | AdamW | AdamW |
| Trainable Params | LLM only | Full (vision + LLM) |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Jingcz/RobustRDP-Pretrain-SFT"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
Related Resources
- Dataset: RxnLabelData
- Annotation Platform: RxnLabel