jordangong/the-stack-v2-smollm3

The Stack v2 — materialized source code

Upstream dataset:

bigcode/the-stack-v2

Exact upstream commit:

e565caa3a78c2423bd374333a472b049eb090e47

Primary source-content endpoint:

https://softwareheritage.s3.amazonaws.com/content/{blob_id}

Configurations

• TypeScript
• Swift
• Ruby
• Rust
• Go
• Shell
• Jupyter_Notebook
• HTML
• Python
• Java
• JavaScript
• C
• C++
• C-Sharp
• PHP
• SQL
• Markdown

Added columns

• content: decoded source content
• download_error: null on successful materialization

An isolated missing or unavailable Software Heritage object does not stop the full materialization. Its original Stack v2 metadata row is retained with content = null.

Failure records are stored under _failures/<language>-shards-<first>-<last>/failures-<index>.jsonl, aligned with the finalized Parquet shard index.

Objects recovered through the Software Heritage archive API after an S3 failure are recorded under:

_recoveries/<language>/

New bucketed recovery records are stored under _recoveries_sharded/.

Reproducibility

All upstream metadata comes from the exact commit above.

Successful source blobs are SHA-1 verified against their blob_id.

Machine-readable source provenance is stored in:

provenance.json

Final layout metadata and cached source layouts are stored under:

_layout/

Individual materializer run configurations are stored under:

_runs/

Per-language or per-partition checkpoints are stored under:

_state/

The finalized Parquet layout is data/<language>-shards-<first>-<last>/train-<index>.parquet, with at most 8000 files per leaf directory. Legacy data/<language>/ and data_sharded/<language>/<partition>/... paths were atomically folded into this single layout after a checksum and checkpoint audit.

Materialization was distributed across multiple workers. Historical per-partition checkpoints and audit records retain their part-<index>-of-<count> tags for reproducibility; finalized Parquet paths no longer include worker partitions.

Known default languages are assigned using both raw bytes and file counts from the language_stats.csv snapshot identified in each run manifest. This balances content-transfer volume and the number of individual Software Heritage HTTP requests while preserving the configured language order within each worker.

Partition layouts are locked under _partitioning/. Do not change a language's partition count after its first run; choose a new destination repo if a different layout is required.

Parquet shards are uploaded in batches of up to 16. Each batch, its advanced checkpoint, and its associated failure/recovery audit files are committed atomically. Hub uploads use 2 concurrent file transfer(s) per worker and finite HTTP timeouts so a stalled transfer can be retried by the worker launcher.