birgermoell/oellm-longctx-tokenized-superlong-512k-1m-2m-v1

OELLM Superlong Long-Context Tokenized 512K/1M/2M v1

This dataset is a superlong-context continuation-training add-on for extending beyond 256K toward 1M-2M context windows.

The design goal is not simply longer packed text. At 1M-2M, the model needs collection-level continuity and explicit pressure to use very old evidence. This artifact therefore mixes real long structured/natural sources with a small multilingual full-span recall component.

Source families:

• RFC Editor plain-text specifications, packed in RFC-number order
• curated technical documentation files from Rust, Django, Kubernetes, Python, and PostgreSQL docs
• multilingual Gutenberg books
• OLMo arXiv full-paper text
• accessible repo-packed code fallback from edward-io/starcoderdata-repo
• synthetic multilingual full-span recall traces generated by scripts/make_2m_curriculum_data.py

The corpus contains 598 packed examples across 512K, 1M, and 2M tiers. The raw pack manifest reports approximately 568M source-side estimated tokens.

Contents

The artifact layout follows openeuro-longctx-megatron-v1:

• bin/: Megatron .bin / .idx files plus tokenization summary
• mix/: weighted Megatron --data-path files
• manifests/: build metadata, checksums, pack/source manifests, recipe, and helper scripts

The staged mix emphasizes 1M/2M tiers:

• RFC specs: 24%
• books: 22%
• arXiv: 15%
• repo-packed code: 15%
• technical docs: 8%
• synthetic recall: 16%

Usage

Download the dataset and use the generated Megatron data path:

python -m longctx.cli artifacts download \
  --repo-id birgermoell/oellm-longctx-tokenized-superlong-512k-1m-2m-v1 \
  --output-dir ./data/oellm-superlong-512k-1m-2m-v1

export SUPERLONG_DATA_PATH="$(cat ./data/oellm-superlong-512k-1m-2m-v1/mix/data_path.args)"

Pass $SUPERLONG_DATA_PATH to Megatron with --data-path, or mix it with the 128K/256K natural and structured datasets as a later-stage continuation component.

Tokenizer

Built with:

• tokenizer type: HuggingFaceTokenizer
• tokenizer model: /home/ubuntu/birger/Megatron-Bridge-utils/tokenizers/openeurollm/tokenizer-256k
• vocab size: 262144

Caveats

This is a practical superlong pilot. The code source is an accessible repo-packed fallback rather than hydrated Stack v2 repositories, because gated Stack v2/SWH access was unavailable on the GPU host. The synthetic recall slice is intentionally included for far-end retrieval pressure and should be kept at a controlled weight.

Source and license note

This is a derived tokenized artifact. Rights and usage constraints follow the upstream datasets and source documents; this dataset card does not grant additional rights beyond those upstream terms.