Storage

Training and evaluation data — metrics, checkpoints, rollouts, trajectories — is saved through a unified storage layer. By default, data is written to local disk. To use cloud storage (GCS, S3, Azure), just change the path to a URI.

For cloud support:

uv pip install 'tinker-cookbook[cloud]'

Training

Pass a cloud URI as log_dir:

# Local (default)
ml_logger = setup_logging(log_dir="/tmp/my_run", config=config)

# GCS
ml_logger = setup_logging(log_dir="gs://bucket/my_run", config=config)

# S3
ml_logger = setup_logging(log_dir="s3://bucket/my_run", config=config)

All training data — metrics, checkpoints, rollouts — is written to the cloud path automatically.

Evaluation

Pass a cloud URI as save_dir:

# Local
config = BenchmarkConfig(save_dir="/tmp/evals/run1")

# GCS
config = BenchmarkConfig(save_dir="gs://bucket/evals/run1")

result = await run_benchmark("gsm8k", client, renderer, config)

Reading Data Back

from tinker_cookbook.stores import TrainingRunStore, storage_from_uri

# Works with local paths and cloud URIs
store = TrainingRunStore(storage_from_uri("gs://bucket/my_run"))

config = store.read_config()
metrics = store.read_metrics()
rollouts = store.read_rollouts(0)
checkpoints = store.read_checkpoints()

Supported Backends

URI	Backend
/local/path	Local filesystem
gs://bucket/prefix	Google Cloud Storage
s3://bucket/prefix	Amazon S3
az://container/prefix	Azure Blob Storage

Any fsspec-supported filesystem can be used.

Flush for cloud backends

Cloud writes are staged locally for performance. Call flush() at checkpoints to ensure data is uploaded:

storage = storage_from_uri("gs://bucket/run")
store = TrainingRunStore(storage)

for step in range(num_steps):
    store.write_metrics({"loss": loss}, step=step)
    if step % save_every == 0:
        storage.flush()

Or use a context manager for automatic flush:

with storage_from_uri("gs://bucket/run") as storage:
    store = TrainingRunStore(storage)
    # ... training ...
# auto-flush on exit