ev-tabpfn 0.1.0

Portable TabPFN evaluation pipeline with baselines, artifacts, reports, CLI, Python API, and MCP tools.

ev-tabpfn

ev-tabpfn is a portable evaluator for TabPFN and tabular baselines. It packages the working Evaluate-TABPFN phase scripts into user-facing concepts:

• data loading
• model execution
• artifacts
• reporting
• batch orchestration
• MCP tools for agents

Install Locally

pip install -e /home/prime/Documents/g3/tab-r1/package

CLI

ev-tabpfn run-single --dataset data.csv --target label --task binary --output outputs/
ev-tabpfn run --config examples/batch_satya_recreation.json
ev-tabpfn aggregate --runs-root outputs/runs --results-dir outputs/results
ev-tabpfn validate --dataset data.csv --target label
ev-tabpfn summarize-run --run-dir outputs/runs/dataset/run_id
ev-tabpfn generate-report --run-dir outputs/runs/dataset/run_id

Python API

from ev_tabpfn import evaluate_dataset, evaluate_batch, aggregate_results, summarize_run

evaluate_dataset(
    dataset_path="data.csv",
    target_column="label",
    task="binary",
    output_root="outputs",
)

evaluate_batch(config_path="config.json")
aggregate_results(output_root="outputs")
summarize_run("outputs/runs/dataset/run_id")

Output Contract

Each dataset run writes:

• predictions/*.csv
• metrics/metrics_summary.csv
• metrics/metrics_summary.json
• metadata/dataset_metadata.json
• metadata/run_config.json
• metadata/model_status.json
• plots/*.png
• logs/*.log

Batch runs additionally write:

• batch_config.resolved.json
• batch_manifest.json
• summary/batch_summary.json
• logs/batch.log