|
--- |
|
license: odbl |
|
--- |
|
Weekly snapshots of Models, Datasets and Papers on the HF Hub |
|
|
|
## Sample code |
|
|
|
To query the dataset to see which snapshots are observable, use e.g.: |
|
```python |
|
import json |
|
|
|
from datasets import load_dataset |
|
from huggingface_hub import HfApi |
|
|
|
REPO_ID = "hfmlsoc/hub_weekly_snapshots" |
|
|
|
hf_api = HfApi() |
|
all_files = hf_api.list_repo_files(repo_id=REPO_ID, repo_type="dataset") |
|
|
|
repo_type_to_snapshots = {} |
|
for repo_fpath in all_files: |
|
if ".parquet" in repo_fpath: |
|
repo_type = repo_fpath.split("/")[0] |
|
repo_type_to_snapshots[repo_type] = repo_type_to_snapshots.get(repo_type, []) + [repo_fpath] |
|
|
|
for repo_type in repo_type_to_snapshots: |
|
repo_type_to_snapshots[repo_type] = sorted(repo_type_to_snapshots[repo_type], key=lambda x:x.split("/")[1]) |
|
|
|
repo_type_to_snapshots |
|
``` |
|
|
|
You can then load a specific snapshot as e.g.: |
|
```python |
|
date = "2025-01-01" |
|
snapshot = load_dataset(REPO_ID, data_files={date.replace("-",""): f"datasets/{date}/datasets.parquet"}) |
|
snapshot |
|
``` |
|
|
|
Returning: |
|
``` |
|
DatasetDict({ |
|
20250101: Dataset({ |
|
features: ['_id', 'id', 'author', 'cardData', 'disabled', 'gated', 'lastModified', 'likes', 'trendingScore', 'private', 'sha', 'description', 'downloads', 'tags', 'createdAt', 'key', 'paperswithcode_id', 'citation'], |
|
num_rows: 276421 |
|
}) |
|
}) |
|
``` |
|
|
|
### Sample analysis of top datasets |
|
|
|
To look at the 10 most liked datasets as of January 1st 2025, you can then run: |
|
```python |
|
[{ |
|
"id": row['id'], |
|
"tags": json.loads(row["cardData"]).get("tags", []), |
|
"tasks": json.loads(row["cardData"]).get("task_categories", []), |
|
"likes": row['likes'], |
|
} for row in snapshot["20250101"].sort("likes", reverse=True).select(range(10))] |
|
``` |
|
|
|
Most of the user-maintained metadata for Hub repositories is stored in the cardData field, which is saved as a JSON-formated string |
|
|