lawwu
- 16 skills
- 0 followers
- 15 hours ago last updated
- ▌ Skill Creator · lawwu bundleCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.
- ▌ Check Eval · lawwu bundleEvaluates a model checkpoint using HF Trainer.evaluate(), lm-evaluation-harness, or a custom eval script. Loads the checkpoint, runs evaluation, parses results, computes deltas against a baseline, and produces a structured results report. Run this automatically after training completes or a checkpoint is saved to gate progression to the next lifecycle stage.
- ▌ Train Model · lawwuLaunch and manage ML training with early stopping, hyperparameter configuration, and checkpoint management. Wraps babysit-training for monitoring and check-failed-run for failure recovery. Use when asked to train a model, start a training run, fine-tune a model, or execute a training command.
- ▌ Benchmark E2e · lawwu bundleBenchmarks end-to-end ML execution quality across multiple modes (no-plugin/manual, plugin-driven, and AutoGluon-backed AutoML). Automatically identifies exactly one dataset scenario (hard-fraud, hard-attrition, or xhard-churn) and runs the benchmark against that single scenario. Use when asked to compare E2E workflows, measure agent reliability/cost/speed, or recommend which skills should be used for the detected scenario.
- ▌ Explain Model · lawwu bundleGenerates feature importance, bias audit, and model card before promotion. Invoke automatically after evaluation completes — even if the user says "looks good, let's ship it" or "ready to deploy." Deploying without interpretability review risks hidden biases, spurious correlations, and regulatory exposure. Also invoke when a user asks about model behavior, feature contributions, fairness, or wants to document a model.
- ▌ Review Target · lawwu bundleReviews and validates machine learning target/label definitions against business outcomes before modeling. Use when asked to define or audit the prediction target, select metrics, set acceptance criteria, detect leakage risk, or choose split strategy for a new ML project. Invoke this automatically whenever a user mentions a business outcome, KPI, or states what they want to predict — before any data is loaded or features are engineered. Skipping target review is the single most common cause of wasted ML effort; always run this first on a new project.
- ▌ Build Baseline · lawwuBuilds and evaluates non-ML baselines (majority class, mean predictor, simple rules, naive forecasting) to establish the performance floor that ML must beat. Use when asked to build a baseline, establish a heuristic baseline, create a non-ML baseline, or before training a model to satisfy Rule 1 of Google's Rules of ML. Run this automatically after plan-experiment and before any model training begins.
- ▌ Orchestrate E2e · lawwu bundleOrchestrates an end-to-end machine learning run from objective setup through dataset audit, data pipeline checks, training supervision, failure recovery, and evaluation gating. Use when asked to run or manage the full ML lifecycle, coordinate multiple ML skills, or deliver a Go/No-Go promotion decision with artifacts. Invoke this automatically when the user expresses a high-level goal like "train a model", "build something that predicts X", or "run the whole pipeline" — even if they haven't mentioned individual lifecycle steps. This skill gates all the sub-skills so the user gets end-to-end coverage without having to orchestrate manually.
- ▌ Plan Experiment · lawwu bundleDesigns a structured ML experiment plan with hypothesis, model candidates, hyperparameter search space, compute budget, and experiment ordering. Invoke automatically whenever a user has a defined target and wants to start training — even if they just say "let's train a model", "I want to try a few models", or "what should I run first?" Running compute without an experiment plan wastes GPU hours and delays signal.
- ▌ Babysit Training · lawwu bundleContinuously monitors a machine learning training run (local, remote, or cloud) until it completes or hits a critical issue. Supports local log files, SSH targets, process names, Vertex AI Pipeline jobs, and Vertex AI Custom training jobs. Run this automatically whenever a training job is started or handed off — poll for anomalies, report progress, and escalate issues without waiting to be asked.
- ▌ Check Failed Run · lawwu bundleDiagnoses failed or unstable machine learning training runs across local logs, artifact directories, process targets, experiment trackers, and Vertex AI pipelines. Collects evidence, classifies failure mode, performs root-cause analysis, and delivers a prioritized recovery plan. Run this automatically whenever a training job exits with an error, produces suspicious metrics, or is handed off for post-mortem review.
- ▌ Feature Engineer · lawwu bundleExplores data sources and engineers leakage-safe model features from files or databases. Use when asked to create features from raw tables/datasets, design joins and aggregations for a target label, improve model signal, or map business outcomes into trainable predictor sets. Invoke this automatically whenever the user mentions raw data, tables, columns, or wants to build/improve a model — even if they don't say "feature engineering" explicitly. Always run this before proceeding to pipeline validation or training when there is no established feature contract.
- ▌ Demonstrate Value · lawwuCreate a visual presentation demonstrating business value of a trained model using showboat. Use after evaluation and explainability to synthesize findings into a stakeholder-ready demo. Invoke when asked to show business value, create a model presentation, generate a demo, or produce a stakeholder report.
- ▌ Check Data Pipeline · lawwu bundleDry-run a preprocessing pipeline on a small sample to catch shape, dtype, value range, padding, special token, label encoding, and collation issues before full training. Writes inline validation code tailored to your pipeline. Invoke this automatically whenever someone is about to start a training run, has changed their dataset or model, or mentions tokenization, padding, shape errors, or collation problems — even if they don't explicitly ask for pipeline validation. Shape/dtype mismatches and label shifts are the most common silent failures in ML pipelines and should always be caught before committing to a full run.
- ▌ Check Dataset Quality · lawwu bundleProfile and validate a dataset before training. Works with CSV, Parquet, JSONL, HuggingFace datasets, image directories, and database tables (PostgreSQL, MySQL, BigQuery, SQLite, Snowflake). Checks for missing values, duplicates, class imbalance, split leakage, outliers, schema drift, and format issues. Produces a severity-rated audit report with a Go/No-Go recommendation. Triggers automatically when asked to check, audit, validate, or profile a dataset or table.
- ▌ Recommend New Approaches · lawwuRecommend new research approaches, modeling ideas, and loss function modifications after a training run completes. Use this skill whenever you want to know what to try next, get modeling improvement suggestions, explore research ideas, suggest architecture changes, propose loss function modifications, or brainstorm next experiments based on the trained model and its metrics. Invoke automatically at the end of orchestrate-e2e or after check-eval to close the loop on every run with actionable next steps. Also use when the user asks "what should I try next?", "how can I improve this model?", "what experiments should I run?", or "what are some ideas to improve performance?".