Skip to workspace

Inspect datasets before you train

Explore nine small, original synthetic datasets with transparent schemas, splits, provenance, and quality checks.

Architecture request routing

Route short requests to reliability, data, or security topics.

text-classification30 rows · 18 / 6 / 6CC0-1.0
nextrun/classification-v1
Provenance, use & limitations
Modality / language
text / English
Features
{"id":"string","split":"string","text":"string","label":"string"}
Source
Original NextRun synthetic teaching fixtures
Updated
2026-09-26
Tags
classification, labels
Intended use
Learn data inspection and pipeline mechanics; not a model-quality benchmark.
Limitations
Only 30 deliberately simple, templated rows. Split IDs are disjoint, but templates recur; replace with independent task families before measuring generalization.
Bias
English-only, hand-designed distributions; no representative population or expert labels.
Sensitive data
No real people, email, customers, or production records. Do not add confidential material to exercises.
Attribution
NextRun Digital, Applied AI Lab synthetic fixtures, v1 (2026).

Inspect the records

id
classification-001
split
train
text
Case 1: Make retries safe after a timeout.
label
reliability
id
classification-002
split
train
text
Case 2: Define the reporting grain before joining.
label
data
id
classification-003
split
train
text
Case 3: Restrict tool credentials to read access.
label
security
id
classification-004
split
train
text
Case 4: Make retries safe after a timeout.
label
reliability
id
classification-005
split
train
text
Case 5: Define the reporting grain before joining.
label
data

18 matches · page 1 of 4

Use the data locally

These examples fetch public teaching fixtures. Package installation and generated files remain in your local environment; external datasets may have separate licenses or access rules.

JSONL · standard library
import json, urllib.request
url = "/static/lab/data/classification-train.jsonl"
rows = [json.loads(line) for line in urllib.request.urlopen(url)]
print(rows[0])
Hugging Face Datasets · requires datasets
from datasets import load_dataset
base = "https://nextrun.io/static/lab/data/classification-"
ds = load_dataset("json", data_files={s: base+s+".jsonl" for s in ["train", "validation", "test"]})
print(ds["train"].features)
Pandas and Parquet · requires pandas + pyarrow
import pandas as pd
df = pd.read_json("/static/lab/data/classification-train.jsonl", lines=True)
df.to_parquet("teaching.parquet", index=False)
print(pd.read_parquet("teaching.parquet").dtypes)
Tokenization and DataLoader · requires datasets, transformers, torch
from datasets import load_dataset
from transformers import AutoTokenizer
from torch.utils.data import DataLoader
ds = load_dataset("json", data_files="https://nextrun.io/static/lab/data/instruction-train.jsonl", split="train")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-2507", revision="cdbee75f17c01a7cc42f958dc650907174af0554")
def encode(row):
    return tok(row["instruction"], truncation=True, padding="max_length", max_length=64)
encoded = ds.map(encode, remove_columns=ds.column_names).with_format("torch")
print(next(iter(DataLoader(encoded, batch_size=2)))["input_ids"].shape)
Split preparation · local dataset with group_id required
# Replace rows with your governed records.
# Keep every source family together, then freeze a versioned manifest.
from sklearn.model_selection import GroupShuffleSplit
first = GroupShuffleSplit(n_splits=1, test_size=0.2, random_state=42)
train_ids, holdout_ids = next(first.split(rows, groups=[r["group_id"] for r in rows]))
# Split holdout groups again into validation and test before model selection.
# Bundled toy rows use predefined splits and are not a generalization benchmark.
Bring this lab to your team Original synthetic fixtures · Not a model-quality benchmark