---
license: odc-by
language:
- en
task_categories:
- text-generation
pretty_name: lilm document-contiguous 16K Longmino pilot data
---

# lilm document-contiguous 16K Longmino pilot data

Curated from `allenai/dolma3_longmino_mix-50B-1025` with `HuggingFaceTB/SmolLM2-135M` by pipeline
`lilm-longctx-16k-v2-cpu-doc-contiguous-longmino`. This is the long-context portion of a recovery pilot;
the intended training mix supplies short-context replay separately.

## Binary format

- `lilm-longctx-16k-train.bin`: training records
- `lilm-longctx-16k-val.bin`: validation records
- dtype: little-endian `uint16`
- record width: 16,385 tokens
- training input/target width: 16,384 tokens
- target long-data budget: 102,000,000 tokens
- each record comes from one document; unrelated documents are never packed together
- consecutive records from the same document overlap by one boundary token
- tokenizer: `HuggingFaceTB/SmolLM2-135M`

Use `numpy.memmap(path, dtype=np.uint16)` and reshape to
`(-1, 16385)`. For each row, use `row[:-1]` as inputs and
`row[1:]` as targets.

`manifest.json`, `sanity_report.json`, and `blocks.jsonl` record provenance,
counts, hashes, and validation results. Sanity status at upload:
`True`.

Upstream dataset license and attribution remain applicable. Repository target:
`glouriousgautam/lilm-longctx-16k-longmino`.