tianchuang/Qwen-3-8B-RHEA-property-predictor
The tianchuang/Qwen-3-8B-RHEA-property-predictor is an 8 billion parameter Qwen-3 model fine-tuned by tianchuang for specialized materials science tasks. It excels at predicting properties and classifying phases of refractory high entropy alloys, including density, hardness, and compressive strength, as well as single solution and intermetallic phase classification. This model leverages a 32K token context length to process detailed alloy compositions and preparation methods for accurate material property predictions. It is specifically optimized for binary classification and regression tasks within materials science research and development.
Loading preview...
Model Overview
The Qwen-3-8B-RHEA-property-predictor is an 8 billion parameter language model developed by tianchuang, specifically fine-tuned for applications in materials science. Its primary function is to predict properties and classify phases of refractory high entropy alloys (RHEA).
Key Capabilities
- Property Prediction: Accurately predicts various mechanical and physical properties of RHEAs, including:
- Density
- Hardness
- Compressive yield strength (at room temperature, 1073K, and 1273K)
- Compressive strain (at room temperature)
- Phase Classification: Determines whether a given RHEA exhibits a single solution phase or an intermetallic phase.
- Specialized Prompting: Utilizes a structured prompt template for materials science queries, incorporating alloy composition and preparation process descriptions.
Performance Metrics
During validation, the model demonstrated strong performance on its specialized tasks:
- Property Prediction (R2 scores): Achieved R2 scores up to 0.886 for density prediction, 0.671 for compressive yield strength at 1273K, and 0.565 for compressive yield strength at 1073K.
- Phase Classification (Accuracy): Reached an accuracy of 0.886 for intermetallic phase classification and 0.858 for single solution phase classification, with F1 scores around 0.840-0.857.
Training Details
The model was fine-tuned using a full fine-tuning approach with CrossEntropyLoss on a specialized dataset of RHEA properties and phases, available at Hugging Face Datasets. It was trained for 4 epochs with a sequence length of 1024 tokens.
Use Cases
This model is ideal for researchers and engineers in materials science who need to quickly assess the properties and phase stability of refractory high entropy alloys based on their composition and processing. It can accelerate materials discovery and design by providing rapid, data-driven predictions.