WikifitaGitHub live67e8de5
pesquisa · kaggle/pokemon_tcg_daily_corpus_training

Pokémon TCG AI Battle — Daily Corpus Orchestration

Resumable daily build-train-clean contract, retained as tooling but restricted to smoke execution during implementation.

Baixar raw

Pokémon TCG AI Battle — Daily Corpus Orchestration

Storage contract

Raw daily replay ZIPs are durable sources. Expanded NPY datasets are large, reproducible, and transient. Checkpoints and sequence state are durable.

retained replay ZIP
        -> deterministic encoded dataset
        -> configured training phase
        -> atomic checkpoint and sequence state
        -> encoded dataset may be removed

The encoded dataset is removed only after checkpoint and state writes succeed. A missing source ZIP prevents cleanup.

Sequence state

The orchestrator records ordered dates, epochs per date, optimizer/scheduler phase policy, output checkpoint, active phase, dataset fingerprint, session sheet hash, and completed dates.

  • interruption before a completed epoch restarts from the recorded base;
  • interruption after an epoch resumes from the rolling checkpoint;
  • a rolling checkpoint is accepted only when its phase identity matches;
  • across daily corpora, the canonical continuation is the latest completed state rather than a validation score from another dataset;
  • an existing checkpoint without matching sequence state is not overwritten by a nominally fresh sequence.

Dataset reuse requires the manifest to match the current would-KO settings, seed, aliases, sides, episode cap, and flush contract.

The underlying replay-to-NPY and prospective-sidecar mechanics are documented in pokemon_tcg_data_pipeline. This page is the orchestration layer: ordered dates, phase state, checkpoint policy, and cleanup boundary.

Operational boundary

The orchestration code is implemented, but no long sequence may be started during the current work. All execution uses configs/smoke.json, real replay archives, and a reduced number of episodes/rows.

Related pages