CycleNER for Materials Science

Quick Start

1. Install dependencies:

pip install torch transformers scikit-learn numpy tqdm
#and
pip install graphviz streamlit langraph langraph-groq

2. Train the model:

python main.py --data_path your_data.json

Key Arguments

• --model_name: Choose t5-small, t5-base, facebook/bart-base, or facebook/bart-large
• --epochs: Number of training epochs (default: 2)
• --batch_size: Training batch size (default: 32)
• --checkpoint_path: Directory to save checkpoints
• --resume_from_checkpoint: Resume from saved checkpoint

How CycleNER Works

CycleNER uses two models working in tandem:

1. S2E Model (Sentence-to-Entity): Extracts entities from sentences
2. E2S Model (Entity-to-Sentence): Generates sentences from entity sequences

Training Process

The model trains through cycle-consistency:

1. S-Cycle:

   • Take sentence → S2E model → predicted entities
   • Train E2S model to reconstruct original sentence from predicted entities

2. E-Cycle:

   • Take entity sequence → E2S model → predicted sentence
   • Train S2E model to reconstruct original entities from predicted sentence

This dual training ensures both models learn complementary representations without requiring labeled data.

Data Processing

The system automatically extracts these entity types from materials science JSON:

• Precursors, targets, solvents
• Temperatures, times, operations
• Quantities and conditions

Entity sequences use format: entity <sep> type <sep> entity <sep> type

Model Comparison

T5 Models: Better for text-to-text generation tasks, explicit task prefixes BART Models: Optimized for text generation and reconstruction tasks

Both architectures work with the same training procedure and data format.