storagestuffnobody/mid-training-40b

Mid-training corpus

39,999,916,670 canonical content tokens in 1,717,205 whole records. The four source configurations preserve text and source metadata. Embedding arrays and chunk payloads are omitted.

Configuration	Records	Canonical tokens	Parquet files
stories	347,887	12,499,999,956	4,136
finepdf	514,578	10,999,999,770	1,758
stack	104,653	7,499,917,120	381
glotcc	750,087	8,999,999,824	541

Counts include every delivered split. FinePDF preserves its source's one-document test split; that inherited split is not presented as an evaluation benchmark. All other configurations have a training split.

Load

from datasets import load_dataset

stories = load_dataset("storagestuffnobody/mid-training-40b", "stories", split="train", streaming=True)
finepdf = load_dataset("storagestuffnobody/mid-training-40b", "finepdf", split="train", streaming=True)
stack = load_dataset("storagestuffnobody/mid-training-40b", "stack", split="train", streaming=True)
glotcc = load_dataset("storagestuffnobody/mid-training-40b", "glotcc", split="train", streaming=True)

Every configuration has a text column, canonical_tokens, a stable mixture_id, and source dataset/revision/file/row fields. Other fields retain source-specific metadata, so the configurations intentionally have different schemas.

For Stack, one record is one repository. Text follows the source serialization, with ### File: <path> headers and blank lines between files. Each files entry retains its original non-content metadata plus text_start, text_end, and content_sha256. Offsets count Python Unicode characters; text[text_start:text_end] recovers that file's exact original contents. Repository order and file order are preserved.

For GlotCC, the original raw_text field is named text. Stories and FinePDF text is preserved exactly. Duplicate HTML bytes and embedding-generation payloads are excluded; source identifiers, document/repository metadata and original scalar content-token counts are retained.

Selection

• Stories: the accepted 12.5B later/longer mask is unchanged. It starts from separately grouped word bands and uses bounded metadata heuristics for author volume, span-aware recency and work length. These weights have not been validated as prose-quality scores.
• FinePDF: a deterministic 11B whole-document sample of the existing semantic-long pool, stratified by its semantic clusters and canonical-token length bands.
• Stack: both parts' existing least-overlap 30% masks, with the later length cut omitted. The two parts were processed separately.
• GlotCC: all 220 clusters were reviewed. The 37 initial removal candidates received independent review, using up to 100 additional random documents each. Twenty-seven clearly bad clusters were excluded; ten candidates with useful content or uncertainty were retained. The resulting 27,368,381,437-token pool was sampled to approximately 9B across all 193 retained groups and token-length bands. NSFW alone was not an exclusion criterion.

GlotCC judgments are based on saved beginning/middle/end excerpts. Some excluded clusters contain rare useful outliers, documented in the review evidence. The clustering run finished at its refinement limit and reports converged=false. This release adds no new cross-source deduplication or downstream training evaluation.

Token totals sum the stored canonical content/body counts from the verified source pipelines. They exclude separately tracked model-input overhead. Different training tokenizers can produce different counts. Records were not truncated or duplicated to force a round budget.

Provenance and verification

manifest.json lists every data file's exact row/token counts, byte size and SHA256. Frozen row-ID masks and source plans are under selection/; quality decisions and their document-ID evidence are under review/. Export verification and code manifests are under verification/ and provenance/.

The local audit checks every output file hash, every selected row/token/cluster, source identities, every text digest and every recoverable Stack file digest. Source metadata equality is also enforced during export. Original upstream datasets remain unchanged.

• stories: storagestuffnobody/stories-embedding
• stack1: storagestuffnobody/stack-v3.1-subsampled-embeddings
• stack2: storagestuffnobody/stack-v3.1-subsampled-part2-embeddings
• finepdf: storagestuffnobody/finepdf-semantic-long
• glotcc: storagestuffnobody/glotcc-over-30k