s0duku's picture
Add sample reading examples
8c9d5ac verified
|
Raw History Blame Contribute Delete
2.76 kB

Reading sample Parquet files

Each samples/train-xxxxx-of-00064.parquet file contains:

sample_id       string
repo_id         string
revision_id     string
kind            dependency | target_context
text            large_string
files           list<struct<path, relation, content_policy, role, language, bytes, part, parts>>
token_count     int32
metadata_json   large_string containing a JSON object

Use streaming for full-dataset inspection. The sample files are independent Parquet files, so the examples work with a single shard or samples/*.parquet.

Command-line inspector

cd stackgo
PYTHONPATH=. python examples/read_samples.py ../samples/train-*.parquet --limit 3

Filter one repository or one sample kind:

PYTHONPATH=. python examples/read_samples.py ../samples \
  --repo owner/project --kind dependency --limit 5

Print full text and JSONL records:

PYTHONPATH=. python examples/read_samples.py ../samples \
  --text-chars 0 --json

PyArrow streaming

import json
import pyarrow.dataset as ds

dataset = ds.dataset("samples", format="parquet")
scanner = dataset.scanner(
    columns=["sample_id", "repo_id", "revision_id", "kind", "text", "files",
             "token_count", "metadata_json"],
    filter=(ds.field("kind") == "dependency") & (ds.field("repo_id") == "owner/project"),
    batch_size=512,
)
for batch in scanner.to_batches():
    for row in batch.to_pylist():
        print(row["sample_id"], row["token_count"])
        print([item["path"] for item in row["files"]])
        print(json.loads(row["metadata_json"]))
        print(row["text"][:500])

Hugging Face datasets

For a local directory or a dataset repository containing samples/*.parquet, use the Parquet loader with streaming enabled so the full dataset is not loaded into memory:

from datasets import load_dataset

samples = load_dataset(
    "parquet",
    data_files="samples/*.parquet",
    split="train",
    streaming=True,
)
for row in samples.take(3):
    print(row["repo_id"], row["kind"], row["token_count"])
    print(row["text"][:500])

If the files are hosted under a dataset repository, point data_files at the repository-relative glob:

samples = load_dataset(
    "parquet",
    data_files="hf://datasets/s0duku/the-stack-go/samples/*.parquet",
    split="train",
    streaming=True,
)

File provenance

files is the authoritative list of files included in the sample. Use relation to distinguish the target from dependencies/repository context and content_policy to distinguish full content from metadata-only entries. metadata_json contains sample-type-specific fields such as the dependency component or target context paths.