CoolFace
Datasetpublic

scholzmx/sepalith

Sepalith dataset Open, R-specialized next-edit-suggestion training data. Private. Layout (read this first) The repo is a projection of the working corpus (a NAS). Everything here is either (a) source data with its license trail, (b) derived synthetic families, or (c) assembled training mixtures. Heavy raw corpora are summarized at ledger/manifest level here; full content lives on the source system (see each corpus entry below). path what it is… See the full description on the dataset page: https://huggingface.co/datasets/scholzmx/sepalith.

sourceHugging Faceupdated 20d agoView on Hugging Face
0likes405downloads
Dataset Card

Sepalith dataset

Open, R-specialized next-edit-suggestion training data. Private.

Layout (read this first)

The repo is a projection of the working corpus (a NAS). Everything here is either (a) source data with its license trail, (b) derived synthetic families, or (c) assembled training mixtures. Heavy raw corpora are summarized at ledger/manifest level here; full content lives on the source system (see each corpus entry below).

pathwhat it is
pretraining/The A2 training package (2026-08-26): token-id block files (int32, 1025×N overlap-1), the 32K tokenizer, the mixture manifest, repack stats + contamination gate — everything run.py all needs on a rented instance (see below)
corpus/cran/The era-1 CRAN corpus: full package shards (packages/), licenses.jsonl, manifest.jsonl
corpus/bioc/Bioconductor corpus: license ledger + manifest (content level: ~0.35-0.55B tokens on source)
corpus/stack_v2/StackOverflow-R corpus: license ledger, measure + dedup reports (1.645B net-new tokens measured)
corpus/stack_v3/GitHub-R tranche: acceptance report (contamination-corrected; 143.5M net-new)
corpus/pwc/Papers-with-code R tranche: license ledger, manifest, measure report (~40M tokens)
families/<family>/<source>.jsonlSynthetic case families — one file per author source. Rows carry base_sample_id (content-hash parent link) + rule/backend tags
mixtures/sft_vX/Assembled train/eval splits (derived; rebuilt from families by experiments/post-processing/assemble_sft_v5.py)
provenance/Per-package provenance records
datasets/Package index + top-level manifest

The pretraining package (pretraining/)

Layout mirrors the training data-root exactly, so the whole package is consumable in two commands from a fresh instance:

hf download scholzmx/sepalith --repo-type dataset \
  --include "pretraining/*" --local-dir /data
python3 run.py all --data-root /data      # verify→repack(skip)→manifest→train→gates
  • —a2_transfers/<stratum>/blocks.npy — transfer strata (english, pythonv2, ccpp, js_ts, sql, julia, matlab, curated python), each with its fetch stats.json (SWH-origin stack-v2 content, permissive-only, sha1-verified)
  • —a2/r/*.npy — the R strata: r_causal (CRAN plain-doc render), r_fim_mix (PSM + random-cursor + noop, the validated mix), `rnoop, sorqa (the Stack v3 R keep set: answer-CODE, ODC-By-1.0, deduped — NOT Q&A prose), bioc` (Bioconductor current R/tests; man/ excluded per the roxygen double-count rule), plus held-out eval slices with byte counts (BPB denominators)
  • —a2/r/stats.json + a2/r/contamination.json — repack accounting and the contamination gate verdict (git-mirror = hard-fail gate, internal eval slices = package-disjoint by construction, warn-only)
  • —datasets/a2_tokenizer_v1/tokenizer.json — the frozen 32K tokenizer (eos <|endoftext|> id 5)
  • —a2_mixture_manifest.json — per-stratum block paths, draw shares, token counts, epochs-equivalent (repetition is surfaced, never silent; sorqa/bioc shares are provisional pending the GO-time re-cut)

Deferred R sources (documented, not in the package): full-CRAN causal beyond astfim's span sample, git/ GitHub-R repos (license audit pending), CRAN Archive (edit-diff), vignettes/prose slice, and the English×R Q&A-prose bridge. Eval-protected assets (the transfer org mirror, R eval holdouts) never enter training data.

Corpus licensing (the standing rule)

Permissive only (MIT/Apache/BSD/GPL-family/CC-BY/CC0; NC/ND excluded and ledgered). Every corpus directory carries its license ledger or report; exclusions are recorded, not silently dropped.

Source labels in families/

glm-5.3 (zai), gemini-3.7-flash (agy), muse-spark-1.2 (opencode GO/free), x-preview (zen), model-id (openrouter), corpus = deterministic/mined (no LLM authorship).