The Agent Skills Directory

Trusted External Resources: The skill fetches protein language models and structural weights from Hugging Face repositories (specifically under the microsoft and facebook organizations). It also utilizes standard content delivery networks like jsDelivr for loading scientific visualization libraries (Mol*).
Remote Dependency Loading: The scoring server's Docker configuration downloads prebuilt Python wheels for specialized Mamba architecture components (causal-conv1d and mamba-ssm) directly from their official GitHub release pages. While these are external binaries, they are standard requirements for the specified model architecture and originate from recognized repositories in the AI research community.
Dynamic Model Execution: The implementation uses the trust_remote_code=True parameter when loading models via the Hugging Face transformers library. This is a common requirement for non-standard model architectures like hybrid Mamba-Transformer models to allow the execution of custom modeling code provided in the model repository. Given the models are authored by the same vendor as the skill, this is consistent with the intended functionality.
In-Memory Progress Tracking: The backend implements a custom in-memory tracker to provide live generation updates to the frontend. This system includes automated garbage collection of stale entries to manage memory usage effectively.
Biosecurity Guardrails: A significant feature of the skill is the inclusion of a 3-layer screening module (sequence_screening.py). It checks protein sequences against a Select Agent toxin reference set (including abrin, ricin, and botulinum neurotoxin) using exact matching, subsequence containment, and percent identity thresholds. This is implemented as a safety filter to prevent the generation of potentially hazardous biological sequences.

msresearch-dayhoff