abcastor/kefir2-casebits

Kefir2 CaseBits — building, not yet a usable corpus

Status: BUILDING / PARTIALLY PUBLISHED. This public repository contains independently checked CaseBits regions. The current verified, gapless raw prefix is named by generations/casebits-nvidia5eaa64-v1/frontier/LATEST.json; resolve its immutable manifest revision and verify the hashes before consumption. Out-of-order data can exist beyond that prefix. Full coverage of the selected 2,100 source regions, global stratification, packed training batches and production training readiness are separate gates and are not asserted by this card.

Source, attribution, and modifications

The intended new generation is derived from NVIDIA Nemotron-ClimbMix, pinned to exact source revision 5eaa64b9c0c85b7f56af01d7dffdb0795816b12b. The source publishes Creative Commons Attribution–NonCommercial 4.0 International (CC BY-NC 4.0) terms, including attribution and non-commercial restrictions. See the source at its frozen revision, the license, and ATTRIBUTION.md. This derivative is designated cc-by-nc-4.0; tokenization is not a mechanism for removing copyright, attribution, or redistribution obligations. Source-specific rights and publication clearance remain important before publishing any source-bearing shards.

The planned generation applies the frozen Kefir1 source filtering/selection rules to the pinned ClimbMix input and then applies the selected new CaseBits learned pattern-mixture representation. This is a new regeneration, not a claim of bitwise identity, identical order/splits, or unchanged text relative to any previous packed corpus. The older small-llm-corpus-100b-v2-dataset bucket is historical provenance/reference only: it is not this generation's raw source. Processing changes, including filtering/selection and binary CaseBits encoding, must be documented in admitted shard manifests when available.

Unpacked canonical representation

Every admitted region carries the following unpacked, little-endian binary components. Each region is bounded; consumers may download one region at a time and read its arrays with standard NumPy memory maps. This is a specialized token-array corpus, not a Parquet/Arrow text dataset.

Component	Meaning
casebits_mixture.ids.u16le	uint16 lower BPE ID per position, including one structural end-of-document ID 0 per document
casebits_mixture.case.u16le	uint16 complete ASCII-uppercase case mask for each ID, including mask 0 for structural EOD
casebits_mixture.offsets.u64le	uint64 document cuts in token indices, D+1 entries, starting at zero and ending at total ID count
raw_bytes.u64le	uint64 original UTF-8 byte lengths for the D documents, excluding EOD
documents.jsonl	Document identity, fixed source filtering scores, thematic cluster, train/validation split and byte/token offsets
receipt.json	Frozen worker source/codec pins, counts and output hashes
admission.json	Independent complete region proof, source attribution and exact file hashes

The lower-BPE vocabulary fingerprint is SHA-256: ab5244cea338b9653ec22b991bae1cee89013ea1dddd71d4c4a2a5c761b5120b. CaseBits is the new learned four-component casing-pattern mixture, not the earlier independent-bit version or an lcmark export. A future consumer must handle the lower IDs and conditional case masks, raw document boundaries, and structural EOD 0 explicitly.

No standard Parquet/Arrow dataset viewer, generic load_dataset(..., streaming=True) interface, or trainer-ready streaming API is promised. These binary streams require a purpose-built, validated reader and revision-pinned metadata.

Reading one admitted region

After downloading the complete file set named by the revision-pinned manifest and verifying sizes and SHA-256, the numeric layout can be inspected without a remote loading script:

from pathlib import Path
import numpy as np

region = Path("downloaded-complete-region")
def read_array(name, dtype):
    path = region / name
    # A fully filtered region legitimately has empty ID/case/raw-byte arrays.
    return (np.memmap(path, dtype=dtype, mode="r")
            if path.stat().st_size else np.empty(0, dtype=dtype))

ids = read_array("casebits_mixture.ids.u16le", "<u2")
cases = read_array("casebits_mixture.case.u16le", "<u2")
offsets = read_array("casebits_mixture.offsets.u64le", "<u8")
raw_bytes = read_array("raw_bytes.u64le", "<u8")
assert len(ids) == len(cases) and len(offsets) == len(raw_bytes) + 1
# Document i occupies ids[offsets[i]:offsets[i+1]], ending in EOD 0 / case mask 0.

These shape checks do not replace the full receipt, semantic or SHA validation. Exact text reconstruction also requires the pinned lower-BPE vocabulary and the complete per-token ASCII uppercase mask. No additional public codec implementation or generic Dataset Viewer is implied.

Readiness, ordering, and limitations

Only shards named by an independently verified, immutable-revision-pinned ready manifest, with a gapless ordered prefix in their declared split and validated file hashes, may be treated as consumable. A file uploaded early or exposed on a public branch is not automatically ready; a missing predecessor, partial set, inconsistent hash, or unverified manifest must prevent consumption. Readiness of the corpus and readiness of any trainer integration are separate gates.

This repository does not assert any current accepted shard count, total documents, future volume, sustained production speed, storage quota, guaranteed availability, or completed coverage. Hugging Face public storage is best-effort and subject to platform limits, not unlimited capacity. Refer to immutable published revisions and future checked manifests for actual evidence.

Attribution and use

Source: NVIDIA, Nemotron-ClimbMix, revision linked above. License: CC BY-NC 4.0 (non-commercial, attribution, indication of changes). Changes: intended frozen-filter regeneration and CaseBits token-derived representation; any additional material changes should be listed in the eventual verified manifests. This project is not affiliated with or endorsed by NVIDIA. See ATTRIBUTION.md.