Blum / scripts /README.md
Renato
Persist market learning and prioritize unseen evidence
70139eb
|
Raw History Blame Contribute Delete
3.66 kB

Script catalog

Documentation · Development · Deployment

These entry points keep stable paths because Docker, CI, tests and trusted jobs call them directly. Run from the repository root with the appropriate dependencies and private environment. The catalog groups them by purpose without changing runtime contracts.

Data inspection and verification

Script Purpose
audit_historical_sources.py Reproduce reviewed packaged-source hashes and acceptance counts
audit_blum_release_dataset.py Inspect a model-release dataset and its gates
build_historical_price_cache.py Build historical price inputs from configured providers; not a prospective track record
verify_ascent_replay.py Reproduce a bounded price-only ASCENT replay offline
benchmark_learning_engine.py Pinned local daily-bar research, persistent-learning workload and frozen chronological validation; no runtime ingestion

Trusted research and model operations

Script Purpose
run_ascent.py Read stored status or explicitly run/resume a bounded experiment
finrlx_runner.py Background runner boundary for the FinRL-X integration
export_blum_model_release.py Prepare a reviewed, redacted model-release dataset
export_blum_training_dataset.py Export reasoning records for inspection; raw exports do not automatically qualify for training
train_blum_challenger.py Optional HF training job code; paid training remains disabled
evaluate_blum_challenger.py Evaluate a candidate against explicit immutable inputs
promote_blum_model.py Explicit authenticated model publication/promotion tooling

The current data quarantine applies before any export is used for new training or promotion. An executable script is not evidence of eligibility or authorization to launch a paid job. Follow the gated model operations guide.

Release and runtime

Script Purpose
check_space_metadata.py Validate the maintained root HF Space card contract
check_release_revision.py Reject stale main revisions or missing required CI jobs
deploy_huggingface.py Fast-forward-only publication of canonical source to HF
verify_deployment.py Check revision, protected content, readiness, public APIs and mutation denial
start.sh Docker entrypoint, embedded PostgreSQL startup and recovery orchestration
postgres_recovery.py Package/restore a checked base-backup and WAL bundle
restore_status_server.py Temporary startup/restore status server
hf_sync_decision.py Retained compatibility classifier for the old HF-to-GitHub sync; no automatic import workflow remains

Recovery and deployment can affect persistent evidence. Use the operations guide, preserve backups and verify readiness before declaring recovery complete. The public observatory does not invoke these administrative operations.