fedops-dataset 0.3.7

Local-first dataset toolkit for multimodal federated learning artifacts (partition/feature/simulation)

fedops-dataset

fedops-dataset is a local-first dataset toolkit for multimodal federated learning (FedMS2-v8 style).

It helps you:

• fetch raw multimodal datasets
• validate dataset roots and expected files
• generate FL artifacts (partition, feature, simulation)
• load per-client records in Python for Simulation and Deployment workflows

Python requirement: >=3.8

Who This Is For

• FL researchers working with multimodal datasets
• engineers running FedMS2-style experiments repeatedly
• teams that want reproducible alpha / ps / pm artifact generation

What This Package Covers

1. Raw data bootstrap:

• fedops-dataset fetch-raw

2. Raw path validation:

• fedops-dataset check-raw-datasets

3. Artifact generation:

• fedops-dataset create-v8

4. Runtime loading API:

• FedOpsLocalDataset

Supported Datasets

• crema_d
• hateful_memes
• ptb-xl

Default clients:

• crema_d: 40
• hateful_memes: 40
• ptb-xl: 20

Install

pip install fedops-dataset

5-Minute Quickstart

1) Define paths

export REPO_ROOT=/path/to/fed-multimodal
export DATA_ROOT=$REPO_ROOT/fed_multimodal/data
export OUTPUT_DIR=$REPO_ROOT/fed_multimodal/output

2) Fetch raw data

# all supported datasets
fedops-dataset fetch-raw --dataset all --data-root "$DATA_ROOT"

Notes:

• hateful_memes default fetch method is direct public git from:
  • https://huggingface.co/datasets/neuralcatcher/hateful_memes

3) Validate raw roots

fedops-dataset check-raw-datasets --data-root "$DATA_ROOT"

4) Generate artifacts (example: hateful_memes)

# dry run first
fedops-dataset create-v8 \
  --dataset hateful_memes \
  --alpha 50 \
  --sample-missing-rate 0.2 \
  --modality-missing-rate 0.2 \
  --repo-root "$REPO_ROOT" \
  --data-root "$DATA_ROOT" \
  --dry-run

# real run
fedops-dataset create-v8 \
  --dataset hateful_memes \
  --alpha 50 \
  --sample-missing-rate 0.2 \
  --modality-missing-rate 0.2 \
  --repo-root "$REPO_ROOT" \
  --data-root "$DATA_ROOT"

5) Load client records in Python

from fedops_dataset import FedOpsLocalDataset

ds = FedOpsLocalDataset(
    dataset="hateful_memes",
    alpha=50,
    sample_missing_rate=0.2,
    modality_missing_rate=0.2,
    repo_root="/path/to/fed-multimodal",
    data_root="/path/to/fed-multimodal/fed_multimodal/data",
)

print(ds.is_prepared())
client0 = ds.client_records(0, use_simulation=True)
print(len(client0))

Parameter Semantics

• alpha: partition heterogeneity control
• sample_missing_rate (ps): sample-level missingness
• modality_missing_rate (pm): modality-level missingness

Token naming examples used in artifact filenames:

• alpha=0.1 -> alpha01
• alpha=5.0 -> alpha50
• alpha=50 -> alpha50

So 5.0 and 50 intentionally resolve to the same alpha token.

CLI Guide

fetch-raw

Use this to prepare raw datasets under your data root.

fedops-dataset fetch-raw --dataset all --data-root "$DATA_ROOT"

Hateful Memes (single recommended command)

fedops-dataset fetch-raw \
  --dataset hateful_memes \
  --data-root "$DATA_ROOT"

This uses the package default public source (neuralcatcher/hateful_memes) via git download.

check-raw-datasets

fedops-dataset check-raw-datasets --data-root "$DATA_ROOT"

Use this before create-v8 to catch path/file issues early.

create-v8

Generates:

• partition JSON
• feature directories
• simulation JSON

fedops-dataset create-v8 \
  --dataset crema_d \
  --alpha 50 \
  --sample-missing-rate 0.2 \
  --modality-missing-rate 0.2 \
  --repo-root "$REPO_ROOT" \
  --data-root "$DATA_ROOT"

Optional controls:

• --no-partition
• --no-features
• --no-simulation
• --num-clients <N>
• --force
• --dry-run

Python API Guide

FedOpsLocalDataset

Direct usage

from fedops_dataset import FedOpsLocalDataset

ds = FedOpsLocalDataset(
    dataset="crema_d",
    alpha=50,
    sample_missing_rate=0.2,
    modality_missing_rate=0.2,
    repo_root="/path/to/fed-multimodal",
    data_root="/path/to/fed-multimodal/fed_multimodal/data",
)

if not ds.is_prepared():
    ds.prepare(dry_run=False)

partition = ds.load_partition()
simulation = ds.load_simulation()
records = ds.client_records(0, use_simulation=True)

Runtime config usage (Flower style)

from fedops_dataset import FedOpsLocalDataset

run_config = {
    "repo-root": "/path/to/fed-multimodal",
    "data-root": "/path/to/fed-multimodal/fed_multimodal/data",
}

# Simulation mode
node_config = {"partition-id": 0, "num-partitions": 40}

ds = FedOpsLocalDataset.from_runtime_config(
    dataset="hateful_memes",
    alpha=50,
    sample_missing_rate=0.2,
    modality_missing_rate=0.2,
    run_config=run_config,
    node_config=node_config,
)

mode = ds.node_mode(node_config)  # simulation
records = ds.client_records_from_node_config(node_config, use_simulation=True)

Simulation vs Deployment

Simulation mode:

• detect with node_config containing partition-id and num-partitions
• use partition-id to resolve client records

Deployment mode:

• if node_config has data-path, it is used as data root
• each node can point to different local storage

Environment Variables (Optional)

export FEDOPS_REPO_ROOT=/path/to/fed-multimodal
export FEDOPS_OUTPUT_DIR=/path/to/fed-multimodal/fed_multimodal/output
export FEDOPS_DATA_ROOT=/path/to/fed-multimodal/fed_multimodal/data
export HATEFUL_MEMES_ROOT=/path/to/fed-multimodal/fed_multimodal/data/hateful_memes

You can use env vars, CLI args, or runtime config keys. No hardcoded path is required.

Troubleshooting

1. partition file not found:

• run create-v8 first
• verify alpha/ps/pm values match existing artifact names

2. hateful_memes fetch fails in git mode:

• ensure git and git-lfs are installed
• rerun the same command after confirming network access

3. Raw dataset validation errors:

• run check-raw-datasets and follow printed hints

4. Alpha confusion (5.0 vs 50):

• both map to token alpha50
• this is intentional for compatibility with existing FedMS2 artifacts

FAQ

1. Do I need to pass --hateful-memes-root always?

• No. By default it resolves to <data-root>/hateful_memes.

2. Can I use this package without Hugging Face uploads?

• Yes. Local-first workflow is the primary mode.

3. Is FedOpsDatasetClient still available?

• Yes. Use it if you also host artifacts in an HF dataset repo.

Maintainer Release

cd fedops_dataset
python -m build
python -m twine check dist/*
python -m twine upload dist/*