jiosephlee/intern-s1-mini-assay-transfer-record-level-v27-carcinogens-general-best
The jiosephlee/intern-s1-mini-assay-transfer-record-level-v27-carcinogens-general-best is an 8 billion parameter model, fine-tuned from the Intern-S1-mini-lm base model, specifically for assay transfer tasks related to carcinogens. It is optimized for predicting carcinogen activity at a record level, demonstrating significantly lower Macro MAE@3 (0.5587) compared to traditional Morgan fingerprint methods. This model excels in chemical property prediction, particularly for carcinogenicity, making it suitable for pharmaceutical and chemical research applications.
Loading preview...
Model Overview
This model, jiosephlee/intern-s1-mini-assay-transfer-record-level-v27-carcinogens-general-best, is an 8 billion parameter checkpoint derived from the Intern-S1-mini-lm base model. It has been specifically fine-tuned for the assay transfer task concerning carcinogens, focusing on predicting activity at a record level. The model was selected at training step 500 based on its joint ID/OOD level-macro MAE@3 performance.
Key Capabilities
- Carcinogen Activity Prediction: Specialized in predicting carcinogen activity with high accuracy.
- Improved Performance: Achieves a Macro MAE@3 of 0.5587, significantly outperforming traditional Morgan fingerprint methods (e.g., vanilla Morgan fingerprint at 0.8142).
- Robust Validation: Selected based on a robust validation metric (
Carcinogens/degree24/overall/knn_regression_id_ood_level_macro_mae_at_3) to ensure reliable performance.
When to Use This Model
- Chemical Property Prediction: Ideal for tasks requiring accurate prediction of chemical properties, specifically carcinogenicity.
- Assay Transfer: Suited for transferring knowledge from existing assays to new, related tasks.
- Research and Development: Valuable for pharmaceutical and chemical research where understanding and predicting carcinogen behavior is critical.
Usage Notes
To load and use this model, standard Transformers APIs should be employed, with trust_remote_code=True enabled to utilize the bundled tokenizer implementation.