ghostcc3/mix-context-post-training-128k

Mix-Context Post-Training Dataset for 128K Context Extension

Overview

Mix-Context Post-Training 128K is a dataset designed specifically for post-training context window extension of pretrained LLMs.

It targets the stage after base pretraining, where a model is adapted to operate over much longer contexts (up to 128K tokens) while preserving short-context behavior. The dataset mixes short- and long-context packed sequences with a controlled length distribution to support:

• Post-training context window extension
• Length generalization / robustness evaluation
• Continued training after positional / RoPE scaling methods

If you use this dataset for post-training, context window extension, or evaluation, please cite this dataset (see Citation).

Sequence Length Distribution and Data Sources

Context Type	Token Length Range	Packed Context Length	Samples	Data Source
Short	64 – 2,048	8K	8,000	FineWeb-Edu (sample/10BT)
Short	2,048 – 4,096	8K	8,000	FineWeb-Edu (sample/10BT)
Short	4,096 – 9,216	8K	16,000	FineWeb-Edu (sample/10BT)
Long	8K – 32K	128K	8,000	RedPajama-Data-1T (arXiv, Wikipedia, Common Crawl)
Long	32K – 64K	128K	8,000	RedPajama-Data-1T (arXiv, Wikipedia, Common Crawl)
Long	64K – 128K	128K	16,000	RedPajama-Data-1T (arXiv, Common Crawl)
Long	128K – 200K	128K	8,000	RedPajama-Data-1T (arXiv, Common Crawl)
Total	—	—	72,000	—

---

Dataset Format

Each example is a packed sequence ready for causal LM training:

• input_ids: token IDs
• position_ids: positional indices aligned to the packed sequence

Note: This dataset does not include raw text. It contains tokenized, packed sequences produced by the preprocessing pipeline.

---

Construction Summary (High-Level)

This dataset is generated by:

1. Downloading public corpora used for short- and long-context content
2. Tokenizing with a specified tokenizer (default in scripts: meta-llama/Meta-Llama-3-8B)
3. Filtering and bucketing by token length
4. Packing sequences to target context windows
5. Concatenating short- and long-context components into the final dataset

Tokenizer

• Tokenizer name/path: meta-llama/Meta-Llama-3-8B
• Each text is encoded with explicit BOS/EOS:
  • BOS + text + EOS
• Length statistics and buckets are tokenizer-dependent

Short-Context Component

• Source: FineWeb-Edu (HuggingFaceFW/fineweb-edu, sample/10BT)
• Bucketed by token length (target sample sizes):
  • 64–2,048: 8,000
  • 2,048–4,096: 8,000
  • 4,096–9,216: 16,000
• Packed to 8K context (short context length)

Long-Context Component

• Source: RedPajama-Data-1T (togethercomputer/RedPajama-Data-1T)
• Splits used:
  • arxiv
  • wikipedia
  • common_crawl (subset used in preprocessing)
• Documents are filtered before tokenization by raw byte length (approx):
  • min: 32 KB
  • max: 800 KB
• After tokenization, long sequences are filtered and bucketed in token ranges:
  • 8K–32K, 32K–64K, 64K–128K, 128K–200K
• Packed to 128K context (long context length)

Packing / Sequence Construction

Packing concatenates tokenized samples sequentially until reaching max_seq_len:

• max_seq_len = 128K
• Short packing context_len = 8K
• Long packing context_len = 128K

---

Citation

If you use this dataset, please cite:

@dataset{wang_chen_mix_context_post_training_128k_2026,
  author       = {Qi Wang and Lizhang Chen},
  title        = {Mix-Context Post-Training Dataset for 128K Context Extension},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/ghostcc3/mix-context-post-training-128k}
}