How to use the PyTDC agent skill for TDC datasets

What the pytdc skill makes the agent do

An agent skill is a folder of instructions and helper scripts that an AI agent reads when a task matches the skill's description. This one covers TDC work: finding which tasks and datasets exist, loading a dataset, creating a train, validation and test split, scoring predictions with standard evaluation metrics, and working with benchmark groups.

The skill asks the agent to follow a fixed order. It first looks up names using package metadata only, which does not download data. It then records the task, dataset, license, expected size, cache folder, split method and random seed. Data is fetched only after you approve that plan. The skill can also audit a local data cache and keep printed output short, showing counts and small previews instead of whole datasets.

Install the pytdc skill in Claude

Download the skill ZIP from this directory page and unzip it. For a personal setup, place the folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder. In the Claude apps you can instead upload the ZIP under Settings, then Capabilities, then Skills.

The skill lists these requirements: uv, CPython 3.11, the PyTDC 1.1.15 Python package, setuptools 80.9.0, network access for dataset and benchmark downloads, and enough disk space for the environment and datasets. Some oracles also need optional service or docking dependencies. The skill's own commands install the pinned packages into an isolated environment, so you do not need to install PyTDC by hand.

• Install uv first, then confirm the skill folder contains a SKILL.md file.
• Expect the environment alone to take hundreds of megabytes before any dataset is downloaded.

Ask for work with natural requests

You do not need to name the skill. The skill is triggered when your request involves TDC therapeutic machine-learning datasets or benchmarks. Name the task area, the dataset if you know it, and what you want to end up with.

Good requests state the goal and leave the download decision to the approval step. The agent should come back with a plan and wait for your go-ahead before fetching anything.

• Which ADME datasets are available in TDC? Show names only.
• List the benchmark groups for ADMET and tell me the license of each.
• Plan a scaffold split for Caco2_Wang with seed 42, but do not download yet.
• Once I approve, load Caco2_Wang and save it under .pytdc-data.
• Score my prediction file against the benchmark metric for this group and print only the summary.

Run the bundled scripts directly

The skill includes scripts for discovery, planning, benchmark evaluation, cache auditing, molecular generation and oracle scoring. Run from the skill directory, a metadata-only discovery command looks like this. Without the --execute flag, the split script only plans and does not download.

Scripts that fetch missing data need --execute. MolGen corpora and supported oracle checkpoints also need --download. Check the reference files in the references folder for details on datasets, oracles and sources.

• scripts/discover_metadata.py lists tasks, datasets, benchmarks, evaluators or oracles.
• scripts/load_and_split_data.py plans or executes a dataset load and split.
• scripts/benchmark_evaluation.py scores predictions against benchmark metrics.
• scripts/cache_audit.py reports on a local data cache.
• scripts/molecular_generation.py and scripts/_common.py support generation and shared helpers.

Understand the safety and network limits

The PyTDC code is MIT licensed, but each dataset has its own terms. Some are Creative Commons licenses, some are non-commercial, and some are listed as not specified. The skill tells the agent to check the dataset's official page and original source before any download, redistribution, publication or commercial use, and to cite both TDC and the original dataset.

Dataset loaders fetch missing data on construction, and some benchmark archives and model checkpoints are large. Remote and docking oracles can send molecular structures to an outside service. PyTDC does not provide a cache quota, eviction policy or dataset-wide checksum, so you manage disk space yourself with the cache audit script.

• Do not run the agent on clinical or medical decisions; these datasets are research resources.
• Give the agent a disk budget and a cache folder before a large download.
• Review the dry-run output before installing packages.
• Check the license on each dataset page before using results commercially.

Tips for better results

Ask for exact dataset names. The skill recommends matching names explicitly rather than relying on fuzzy matching, which can silently pick the wrong dataset. If you are unsure of a name, ask for a discovery listing first.

Keep a fixed random seed and a named split method in every request so that another person can reproduce your split.

• Start with a metadata-only listing, then plan, then approve.
• Keep cache folders separate for each project.
• Ask for previews and counts rather than full files when you review results.