---
pretty_name: anneal_pt_v3 clean v2 — balanced 7,168 shards
task_categories:
  - text-generation
size_categories:
  - n>1B
tags:
  - pretraining
  - parquet
  - cleaned
  - distributed-training
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.parquet
---

# anneal_pt_v3 clean v2 — balanced 7,168 shards

This is the cleaned and deterministically re-sharded training split of
`anneal_pt_v3`. It contains **2,103,321,326 rows** in **7,168 Parquet files**
(3,376,320,935,167 stored bytes).

The shard count is `8 × lcm(128, 448)`, so it divides evenly for both intended
data-parallel configurations:

| Data ranks | Shards per rank | Maximum relative load deviation |
| ---: | ---: | ---: |
| 128 | 56 | 1.4942% |
| 448 | 16 | 0.3888% |

The files are ordered by a fixed-capacity LPT assignment based on uncompressed
`text` bytes. Consumers that assign ordered, equally sized contiguous file
ranges to ranks therefore receive the measured balance above.

## Data fields

Cleaning and re-sharding preserve the source Parquet schema. The primary fields
are `id`, `text`, `source`, and `url`. This repository exposes one `train`
split through `data/*.parquet`.

## Cleaning

The conservative v2 policy quarantined 9,614 of 2,103,330,940 source rows
(approximately 0.000457%). Quarantined rows are **not included** in this public
dataset repository. A row was quarantined when it matched one or more of:

- hard maximum size over 64 MiB;
- repeated-line content in a long document;
- a very large document concentrated in one line;
- excessive control characters;
- an identical-character run longer than 10,000 characters;
- a high Unicode replacement-character ratio;
- whitespace-heavy, low-alphanumeric padding;
- a large encoded/base64-like blob.

The policy thresholds and aggregate reason counts are included under
`metadata/`. Reason counts can overlap because one row may match multiple
rules. No surviving row content was rewritten: cleaning selected complete rows,
and re-sharding only reassigned complete Parquet row groups to output files.

## Validation

- 2,048/2,048 source-cleaning validations passed.
- Source row conservation passed: kept + quarantined = source.
- The re-sharded output contains exactly 7,168 expected data objects.
- Global row count, object set, and rank-load validation passed.

See `metadata/full-run-summary-v2.json`,
`metadata/reshard-summary-7168.json`, and
`metadata/reshard-plan-7168.json` for machine-readable details. The completed
S3-to-Hugging-Face object and byte reconciliation is recorded in
`metadata/hf-upload-summary.json`.

## Usage

```python
from datasets import load_dataset

dataset = load_dataset(
    "rides21/anneal_pt_v3_clean_v2_reshard_7168",
    streaming=True,
)
train = dataset["train"]
```

For large distributed jobs, use the repository file order from
`metadata/reshard-plan-7168.json` rather than relying on an arbitrary object
listing order.

## Content notice

This is a large web-derived corpus. Automated anomaly filtering is not a
substitute for downstream safety, privacy, licensing, or domain-specific
review. Users should apply controls appropriate to their use case.