README.md

---
pretty_name: Dataset Distillation Collection
tags:
- dataset-distillation
- computer-vision
- parquet
configs:
- config_name: images
  default: true
  data_files:
  - split: train
    path: data/images/**/*.parquet
- config_name: teachers
  data_files:
  - split: train
    path: data/teachers/*.parquet
- config_name: manifest
  data_files:
  - split: train
    path: data/manifest/*.parquet
dataset_info:
- config_name: images
  features:
  - name: method
    dtype: string
  - name: dataset
    dtype: string
  - name: ipc
    dtype: int32
  - name: class_id
    dtype: int32
  - name: image_id
    dtype: int32
  - name: source_path
    dtype: string
  - name: extension
    dtype: string
  - name: image
    dtype: image
  - name: sample_weight
    dtype: float64
  - name: sha256
    dtype: string
  - name: byte_size
    dtype: int64
  - name: source_archive
    dtype: string
  splits:
  - name: train
    num_examples: 481090
- config_name: teachers
  features:
  - name: provider
    dtype: string
  - name: dataset
    dtype: string
  - name: architecture
    dtype: string
  - name: used_by
    sequence: string
  - name: source_path
    dtype: string
  - name: checkpoint
    dtype: binary
  - name: sha256
    dtype: string
  - name: byte_size
    dtype: int64
  - name: source_archive
    dtype: string
  splits:
  - name: train
    num_examples: 6
- config_name: manifest
  features:
  - name: asset_type
    dtype: string
  - name: relative_path
    dtype: string
  - name: method_or_provider
    dtype: string
  - name: row_count
    dtype: int64
  - name: payload_bytes
    dtype: int64
  - name: parquet_bytes
    dtype: int64
  - name: sha256
    dtype: string
  splits:
  - name: train
    num_examples: 26
---

# Dataset Distillation Collection

This repository is a byte-preserving Parquet conversion of the assets audited in:

- `synthetic_imagefolders.tar.gz`
- `other_pretrained_models.tar`

It contains 481,090 synthetic images from 13 dataset-distillation methods and six ResNet-18 teacher checkpoints that were physically present in the teacher archive. The original compressed image bytes and checkpoint bytes are stored directly in Parquet binary columns; no image re-encoding or checkpoint rewriting was performed.

## Configurations

### `images`

One row per physical image. Important columns:

- `method`, `dataset`, `ipc`, `class_id`, and `image_id` identify the experiment cell and sample.
- `image` is the Hugging Face image feature backed by the original JPEG/PNG bytes.
- `sample_weight` is populated for all 73,800 WMDD rows and null for other methods.
- `sha256` and `byte_size` validate the original compressed image payload.
- `source_path` and `source_archive` preserve provenance.

`ipc` is the nominal IPC recorded in the source folder. FreD and NCFM deliberately contain more physical images than `classes × ipc`; every physical image is retained.

### `teachers`

One row per teacher checkpoint. The `checkpoint` column contains the original `.pth` bytes. The six included checkpoints are:

- CVDD ResNet-18 for CIFAR-10 and CIFAR-100
- G-VBSM ResNet-18 for CIFAR-10 and CIFAR-100
- WMDD ResNet-18 for TinyImageNet-200 and ImageNette

The source archive did not contain the following required assets, so this repository does not fabricate or substitute them:

- `other_pretrained_models/GVBSM/tiny/ResNet18/squeeze_ResNet18.pth`
- `other_pretrained_models/SRe2L/cifar10/ckpt.pth`
- `other_pretrained_models/SRe2L/cifar100/ckpt.pth`
- torchvision's official ImageNet ResNet-18 checkpoint (a runtime dependency in the project)

### `manifest`

One row per uploaded image or teacher Parquet shard, including row counts, payload sizes, Parquet sizes, and SHA-256 digests.

## Usage

```python
from datasets import load_dataset

# Stream images without downloading the whole collection.
images = load_dataset(
    "Passenger555/DatasetDistillationCollection",
    "images",
    split="train",
    streaming=True,
)
first_image = next(iter(images))

# Restore a teacher checkpoint byte-for-byte.
teachers = load_dataset(
    "Passenger555/DatasetDistillationCollection",
    "teachers",
    split="train",
    streaming=True,
)
teacher = next(iter(teachers))
with open("teacher.pth", "wb") as handle:
    handle.write(teacher["checkpoint"])
```

## Validation

The conversion was independently read back before upload. Validation covered all 481,090 image payloads and all six checkpoint payloads:

- recomputed SHA-256 matched every Parquet row;
- image row counts matched all 119 audited method/dataset/IPC groups;
- one image from each group decoded successfully;
- all WMDD sample weights were present and all non-WMDD weights were null;
- every shard matched the SHA-256 and file size recorded in the manifest.

See `DD_ASSET_COMPLETENESS_CHECKLIST.md` for the complete coverage matrix and `conversion_summary.json` for machine-readable counts and checksums.