Add Nemotron Pretraining Code v1 to pretraining dataset registry

[dlwh]

This adds wiring for nvidia/Nemotron-Pretraining-Code-v1 in the pretraining dataset registry so it can be downloaded and tokenized through the standard experiments/pretraining_datasets CLI.

The new dataset entry targets the Synthetic-Code config (the only config with a text field) and excludes the metadata-only config by restricting HF download globs to Synthetic-Code/*.parquet plus dataset docs.

Changes:

• add downloads["nemotron_pretraining_code_v1"] in experiments/pretraining_datasets/simple.py
• add tokenized["nemotron_pretraining_code_v1"] in experiments/pretraining_datasets/simple.py
• register nemotron_pretraining_code_v1 in experiments/pretraining_datasets/__init__.py

Validation:

• uv run experiments/pretraining_datasets/main.py list
• ./infra/pre-commit.py experiments/pretraining_datasets/simple.py experiments/pretraining_datasets/__init__.py
• uv run python - <<'PY' ... get_steps(['nemotron_pretraining_code_v1'], download=True, tokenize=True) ... PY

[claude]

Claude finished @dlwh's task in 43s —— View job

---

Review: Add Nemotron Pretraining Code v1

Adds download and tokenize wiring for nvidia/Nemotron-Pretraining-Code-v1 to the pretraining dataset registry. The change follows existing patterns correctly and fulfills its stated objectives.

• Gather context and read changed files
• Verify patterns against existing dataset entries
• Check hf_urls_glob usage and text_key default

Findings: No issues found.

The implementation is consistent with other simple dataset entries:

• hf_urls_glob restricts downloads to Synthetic-Code/*.parquet plus docs, which correctly avoids the metadata-only config.
• .cd("Synthetic-Code") scopes the tokenizer to the right subdirectory.
• The default TextLmDatasetFormat(text_key="text") matches the Synthetic-Code schema.
• The __init__.py registry entry follows the exact same structure as the surrounding entries.

LGTM.