mirror of
https://github.com/wandb/skills.git
synced 2026-09-19 09:00:04 +08:00
Zubin Aysola
a8aaad7645
Instead of hardcoding uv as the package manager, the skill now instructs agents to detect the user's Python environment first and write the finding into the skill itself. Falls back to uv only when the environment cannot be determined. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
skills
-yellow){kind=icon}
-brightgreen){kind=icon}
Skills to guide Claude Code, Codex, and other coding agents on using the Weights & Biases AI developer platform to train models and build agents.
For model training
- Log metrics and rich media during model training and fine-tuning
- Track model training experiments
- Analyze runs and experiment results to understand how the model is learning
- Tune hyperparameters
For agent building
- Trace agentic AI applications
- Analyze traces and classify them into failure modes
- Evaluate models with labeled datasets
- Run online evaluations for production monitoring
Getting Started
npx skills add wandb/skills
Then set your W&B API key:
export WANDB_API_KEY=<your-key>
npx skillsis a utility for installing skills into major coding agent CLIs. Use--globalto install for all projects, or--agent <name>to target a specific agent. See the npx skills docs for more details.
Available Skills
| Skill | Description | Status |
|---|---|---|
wandb-primary |
Comprehensive primary skill for agents working with Weights & Biases. Covers both the W&B and Weave SDK | claude-code: 32/35 (91%) |
Benchmarks
We maintain a growing internal benchmark suite that evaluates each skill across coding agents and task categories. Skills are evaluated automatically on every merge to main.
| Category | Tasks | Claude Code (sonnet4.6) |
Codex (gpt-5.3-codex) |
|---|---|---|---|
| Weave analysis | 26 | 97%* | 63%* |
| Weave tooling | 11 | 95%* | 83%* |
| Model training | 8 | 90%* | 85%* |
| LLM finetuning & RL analysis | 14 | 72%* | 86%* |
| Failure & outlier detection | 8 | 86%* | 63%* |
*Pass rates are +/- 3%. Many tasks span multiple categories.
Contributing
See CONTRIBUTING.md.