The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill interacts with untrusted external data by loading molecular datasets from various file formats, creating a potential surface for indirect prompt injection if data contains adversarial content.
Ingestion points: The skill utilizes dc.data.CSVLoader, dc.data.SDFLoader, dc.data.FASTALoader, dc.data.ImageLoader, and dc.data.JsonLoader in SKILL.md and referenced scripts (e.g., scripts/graph_neural_network.py and scripts/predict_solubility.py) to process user-supplied files.
Boundary markers: Absent; the skill processes molecular structures (SMILES strings, SDF data) directly as model inputs without explicit natural language delimiters.
Capability inventory: The skill is capable of reading local files, performing network operations to download datasets and models, and executing machine learning training and prediction logic via model.fit() and model.predict().
Sanitization: The skill relies on DeepChem's featurizers (like CircularFingerprint or MolGraphConvFeaturizer) which treat inputs as chemical entities, effectively limiting the scope for natural language instruction processing from the data.
[EXTERNAL_DOWNLOADS]: The skill references and fetches resources from well-known and trusted technology services for its primary operations.
Fetches standardized benchmark datasets (MoleculeNet) using dc.molnet.load_* functions in SKILL.md, scripts/graph_neural_network.py, and scripts/transfer_learning.py.
Downloads pretrained molecular models (e.g., ChemBERTa, MoLFormer) from Hugging Face's official repository using dc.models.HuggingFaceModel in scripts/transfer_learning.py.

deepchem