TA

tabpfn-feature-encoder-runner

Configures and runs TabPFN training pipelines.

Install

mkdir -p .claude/skills/tabpfn-feature-encoder-runner && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/13566" && unzip -o skill.zip -d .claude/skills/tabpfn-feature-encoder-runner && rm skill.zip

Installs to .claude/skills/tabpfn-feature-encoder-runner

Activation

This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.

Use when running, configuring, validating, or documenting the tabpfn-feature-encoder training repo, including conda setup, runner scripts, output artifacts, tests, and Git hygiene.
180 chars✓ has a “when” trigger
Advanced

Key capabilities

  • Run the `tabpfn-feature-encoder` training repository.
  • Configure TabPFN experiments using YAML files.
  • Validate output artifacts and Git hygiene.
  • Execute performance evaluations for source-task, CP, and open-data transfers.
  • Manage the conda environment for the feature encoder.

How it works

This skill manages the `tabpfn-feature-encoder` training workflow by running scripts for training, transfer evaluations, and plotting, while ensuring the correct conda environment setup and output artifact generation.

Inputs & outputs

You give it
A request to run, configure, validate, or document the `tabpfn-feature-encoder` training repo.
You get back
Trained models, evaluation metrics (e.g., `metrics.json`, `epoch_metrics.csv`), generalization results, and comparison plots.

When to use tabpfn-feature-encoder-runner

  • Running encoder training
  • Configuring TabPFN experiments
  • Executing performance evaluations

About this skill

TabPFN Feature Encoder Runner

Repo Basics

  • Repo root: tabpfn-feature-encoder.
  • Main 12-class source residual config: configs/source_residual_mlp.yaml.
  • Particle GNN config: configs/source_gnn.yaml.
  • Particle transformer config: configs/source_transformer.yaml.
  • Full workflow launcher: bash scripts/run_full_workflow.sh.
  • Dispatcher spelling: bash scripts/run full workflow.
  • Launcher: bash scripts/run_source_encoder.sh.
  • Source transfer rerun: bash scripts/run_source_transfer.sh.
  • CP transfer rerun: bash scripts/run_cp_transfer.sh.
  • Open-data transfer rerun: bash scripts/run_gamgam_transfer.sh.
  • Context comparison plots: bash scripts/plot_context_comparison.sh.
  • Test runner: bash scripts/run_tests.sh.
  • Package CLI: tabpfn-encoder-train train --config configs/source_residual_mlp.yaml.
  • Output dir is configured by output_dir.

Environment

Use the existing conda env:

conda activate tabpfn
python -m pip install -e ".[train,atlas,plots]"

The runner falls back to conda run --no-capture-output -n tabpfn if the console script is not on PATH.

Runner Behavior

scripts/run_full_workflow.sh:

  • Runs configs/source_residual_mlp.yaml, configs/source_gnn.yaml, and configs/source_transformer.yaml by default.
  • For each config, trains the 12-class source encoder and then runs source-task, CP even/odd, and open-data transfer evaluations.
  • Runs configs in parallel by default when multiple GPUs are visible, with one config per GPU.
  • Streams per-config logs to the terminal and writes full logs to runs/workflow_logs/<timestamp>/.
  • Set TABPFN_WORKFLOW_STREAM_LOGS=0 to disable live log streaming.
  • Select GPUs with TABPFN_WORKFLOW_GPUS=0,1,2,3 bash scripts/run_full_workflow.sh.
  • Force sequential execution with TABPFN_WORKFLOW_PARALLEL=0 bash scripts/run_full_workflow.sh.
  • Runs context comparison plotting at the end unless TABPFN_WORKFLOW_PLOT=0 is set.
  • Accepts optional config paths to restrict the workflow: bash scripts/run_full_workflow.sh configs/source_residual_mlp.yaml.

scripts/run_source_encoder.sh:

  • Sets TABPFN_MODEL_CACHE_DIR to $SCRATCH/tabpfn_model_cache unless already set.
  • Sets PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True unless already set.
  • Reuses an existing checkpoint from ~/.cache/tabpfn when available.
  • Accepts an optional config path: bash scripts/run_source_encoder.sh configs/other.yaml.
  • Runs the GNN with: bash scripts/run_source_encoder.sh configs/source_gnn.yaml.
  • Runs the transformer with: bash scripts/run_source_encoder.sh configs/source_transformer.yaml.
  • Reruns source-task transfer from a checkpoint with: bash scripts/run_source_transfer.sh.
  • Reruns CP even/odd transfer from a checkpoint with: bash scripts/run_cp_transfer.sh.
  • Reruns open-data transfer from a checkpoint with: bash scripts/run_gamgam_transfer.sh.
  • Plots encoder comparison PDFs with: bash scripts/plot_context_comparison.sh.

Model Layout

  • Keep distinct encoder definitions in separate files under src/tabpfn_feature_encoder/models/.
  • Current modules: mlp.py, feature_gate.py, feature_mixer.py, gnn.py, transformer.py.
  • Keep model selection in models/factory.py.
  • Keep PyTorch import helpers in models/torch_utils.py.
  • Leave models/encoders.py as a compatibility re-export layer, not the place for new model logic.

Source Training Expectations

  • Training uses frozen TabPFN support/query episodes; only encoder weights are optimized.
  • For the 12-class source task, the trainer uses binary ECOC by default: encoder.tabpfn_max_classes: 2, encoder.many_class_redundancy: 4.
  • Default source configs use output_dim: 72, learning_rate: 0.0002, grad_clip_norm: 1.0, and validation_episodes: 8.
  • Source embeddings are detached before fitting the TabPFN support prompt by default (detach_support_gradients: true) so the encoder learns from query gradients without differentiating through prompt construction.
  • Validation is episodic and rotates through validation support/query contexts using the same 50/50 support/query split as training.
  • Epoch logs should include grad_norm_mean, grad_norm_max, and skipped_nonfinite_updates.

Validation Commands

Prefer these before finalizing code changes:

bash scripts/run_tests.sh

Pass pytest selectors through for focused checks, for example:

bash scripts/run_tests.sh tests/test_config.py

Clean generated Python/cache files before committing:

find . -type d -name __pycache__ -prune -exec rm -rf {} +
find src -maxdepth 2 -type d -name '*.egg-info' -prune -exec rm -rf {} +

Artifacts

Training saves:

  • metrics.json
  • training_summary.json
  • epoch_metrics.csv
  • encoder_classifier.pkl
  • run_metadata.json
  • source_generalization/source_12_class_generalization_metrics.json
  • source_generalization/source_12_class_generalization_context_scan_metrics.csv
  • source_generalization/source_12_class_generalization_context_scan_roc_auc.png
  • source_generalization/source_12_class_generalization_baseline_proba.npy
  • source_generalization/source_12_class_generalization_frozen_encoder_proba.npy
  • cp_generalization/cp_even_odd_generalization_metrics.json
  • cp_generalization/cp_even_odd_generalization_context_scan_metrics.csv
  • cp_generalization/cp_even_odd_generalization_context_scan_roc_auc.png
  • cp_generalization/cp_even_odd_generalization_baseline_proba.npy
  • cp_generalization/cp_even_odd_generalization_frozen_encoder_proba.npy
  • open_data_generalization_metrics.json in transfer.output_dir
  • open_data_generalization_context_scan_metrics.csv in transfer.output_dir
  • open_data_generalization_context_scan_roc_auc.png in transfer.output_dir
  • open_data_generalization_baseline_proba.npy in transfer.output_dir
  • open_data_generalization_frozen_encoder_proba.npy in transfer.output_dir
  • context_scan_comparison/*_roc_auc_comparison.pdf
  • context_scan_comparison/*_accuracy_comparison.pdf

Terminal metrics print to three decimals; CSV/JSON keep full precision.

When not to use it

  • When the task is not related to running, configuring, validating, or documenting the `tabpfn-feature-encoder` training repo.
  • When the task involves model logic outside of `src/tabpfn_feature_encoder/models/`.

Prerequisites

condatabpfn conda environment

Limitations

  • This skill is specific to the `tabpfn-feature-encoder` training repository.
  • It relies on predefined configuration files and runner scripts.
  • It requires a `tabpfn` conda environment to be active.

How it compares

This skill orchestrates the entire training and evaluation workflow for the feature encoder, including environment setup and artifact management, providing a standardized and reproducible process compared to manual execution.

Compared to similar skills

tabpfn-feature-encoder-runner side by side with the closest alternatives in the catalog.

SkillInstallsUpdatedSafetyDifficulty
tabpfn-feature-encoder-runner (this skill)03moReviewAdvanced
llama-cpp218moReviewIntermediate
langchain268moReviewIntermediate
unsloth158moNo flagsIntermediate

Try saying

Example prompts that trigger this skill in your AI assistant.

You might also like

llama-cpp

zechenzhangAGI

Runs LLM inference on CPU, Apple Silicon, and consumer GPUs without NVIDIA hardware. Use for edge deployment, M1/M2/M3 Macs, AMD/Intel GPUs, or when CUDA is unavailable. Supports GGUF quantization (1.5-8 bit) for reduced memory and 4-10× speedup vs PyTorch on CPU.

21471

langchain

zechenzhangAGI

Framework for building LLM-powered applications with agents, chains, and RAG. Supports multiple providers (OpenAI, Anthropic, Google), 500+ integrations, ReAct agents, tool calling, memory management, and vector store retrieval. Use for building chatbots, question-answering systems, autonomous agents, or RAG applications. Best for rapid prototyping and production deployments.

26138

unsloth

zechenzhangAGI

Expert guidance for fast fine-tuning with Unsloth - 2-5x faster training, 50-80% less memory, LoRA/QLoRA optimization

15117

llama-factory

zechenzhangAGI

Expert guidance for fine-tuning LLMs with LLaMA-Factory - WebUI no-code, 100+ models, 2/3/4/5/6/8-bit QLoRA, multimodal support

15112

llava

zechenzhangAGI

Large Language and Vision Assistant. Enables visual instruction tuning and image-based conversations. Combines CLIP vision encoder with Vicuna/LLaMA language models. Supports multi-turn image chat, visual question answering, and instruction following. Use for vision-language chatbots or image understanding tasks. Best for conversational image analysis.

7117

cocoindex

cocoindex-io

Comprehensive toolkit for developing with the CocoIndex library. Use when users need to create data transformation pipelines (flows), write custom functions, or operate flows via CLI or API. Covers building ETL workflows for AI data processing, including embedding documents into vector databases, building knowledge graphs, creating search indexes, or processing data streams with incremental updates.

6116

Search skills

Search the agent skills registry