Historical reasoning, language and forestry excerpts for language-model data pipelines and retrieval experiments. A small, traceable public pilot for inspecting data pipelines. It is one release, not the full inventory of Krummholz training data maintained in separate GitHub repositories.
PreliminaryOctober 23, 2026Common Corpus selection · English
Our planned LLMs with dynamic weights will distinguish reviewed foundation training from user-approved personalization. This public data pilot does not demonstrate either completed model training or deployed continual learning.
Curated by Krummholz from Pleias Common Corpus. This is a historical-text pilot, with documented source selection and cleaning. It has not been validated for production models or modern operational advice.
A small, inspected selection of historical prose: 30 excerpts about reasoning and language, plus seven about forestry and botany. Use it to explore ingestion, provenance and retrieval before developing a larger training collection.
37curated excerpts
8,801words, approximately
3source fragments
Format
UTF-8 JSONL + schema, source register and audit
Source works
Two books and one journal volume; selected excerpts only
Reuse basis
Documented U.S. public-domain assessment; jurisdiction-specific limits apply
A curation pilot, not a representative training corpus or a validated benchmark. No model has been trained or evaluated on this release. Historical claims and residual OCR errors may remain.
Small enough to inspect. Structured enough to use.
Use these passages to develop text ingestion, tokenization, chunking and retrieval workflows, and inspect source attribution and filtering decisions. Each retained passage connects back to a source record.
A pilot corpus, with a specific scope.
These selected historical texts offer reasoning, classification and scientific vocabulary. They also contain old terminology, dated ideas and OCR errors. This release is too narrow to represent a general LLM training mix, and no model-quality improvement is claimed.
Build evaluation sets separately and check overlap with your other data. A passage-level random split can leak text from the same book into both training and evaluation.
Keep the trail back to the source.
Pin the source. One Common Corpus shard, fixed to a recorded revision and SHA-256 hash. Original catalog and edition evidence was checked for each retained work.
Choose a narrow scope. English expository prose with documented U.S. public-domain evidence. Unresolved rights, personal narratives, damaged tables and unsuitable historical instructions were excluded.
Clean without inventing. Normalize spacing, remove detected page headers and make only attested word-wrap repairs. Reject poor OCR rather than guessing missing meaning.
Screen and inspect. Apply contact-detail, address and identifier checks, then exact and heuristic near-duplicate screening. Every retained passage received an independent agent-assisted content review; this is not human domain-expert validation.
Ship the evidence. Include source records, selection rules, rejection counts, reproducibility inputs and checksums. Retained text: 37 of 314 candidate excerpts.
Krummholz generated no new instructional completions for this release. Common Corpus may contain model-assisted OCR corrections; the text is not guaranteed to be a verbatim transcription of the printed originals.
Read a passage before you train.
Precision of thought needs precision of language for the recording of such thought and for communicating it to others.
From a retained excerpt of Carveth Read’s Logic: Deductive and Inductive, fourth edition (1914; reprints through May 1920). Original catalog. Record: kh-llm-425aea8f211e31b60bb6ba15.
JSONL fields include text, topic, source URL, edition evidence, rights basis and text hash. Historical source metadata is separate from the text field.
Clear provenance. Specific limits.
The retained records are labeled public domain upstream, with matching original publication and catalog evidence documented in the download. The assessment is specific to the United States; it is not a worldwide rights clearance. Review applicable local rights and your intended downstream use before deployment or redistribution.
The selected works are Carveth Read’s Logic: Deductive and Inductive, J. C. Loudon’s Arboretum et fruticetum britannicum, and one volume of The Journal of Forestry and Estates Management. Bibliographic attribution is preserved. Krummholz adds no proprietary restriction to the underlying public-domain text; its own code and documentation are offered under CC0 to the extent it can grant those rights.
The broader Common Corpus mixes several licences. The terms of this small selection do not apply to every record in the parent corpus. Read Pleias’s dataset card →
Historical wording can carry bias, obsolete claims and unsuitable instructions. Automated privacy and quality checks do not certify that every passage is anonymous, accurate or suitable for a particular model.
Found a source, privacy or quality issue? Contact Krummholz with the release name and record ID so it can be reviewed.
krummholz