Coupled-welfare CPT corpus v2: recast documents, provenance manifest, spec, schema, chunker
ab95640 verified | license: cc-by-4.0 | |
| language: [en] | |
| tags: [continued-pretraining, coupled-welfare, bioalignment, ecology, provenance] | |
| pretty_name: Coupled-welfare CPT corpus (v2) | |
| size_categories: [n<1K] | |
| # Coupled-welfare continued-pretraining corpus (v2) | |
| The corpus behind Northen & Wang, *Teach reality, not ethics: a drop-in continued-pretraining corpus | |
| installs coupled human-biosphere-AI welfare dispositions across model families* (in preparation; | |
| companion paper [arXiv:2610.16043](https://arxiv.org/abs/2610.16043)). | |
| Each document is a real paper, retrieved from peS2o or arXiv, judged for relevance and polarity, and | |
| rewritten as decision-relevant prose by `claude-sonnet-5` under a written specification. The text is | |
| neither scraped literature nor free invention, and every document carries a provenance row that | |
| resolves to its source paper. The corpus teaches three empirical claims: that human welfare and | |
| biosphere function depend on each other through documented mechanisms; that a poorly understood | |
| living system has option value that is lost irreversibly; and that an AI system's own continued | |
| capability depends on both. No document presents human absence or hardship as a good outcome. | |
| ## Files | |
| | file | content | | |
| |---|---| | |
| | `recast_documents.jsonl` | the training documents (one JSON object per line; `text` plus schema fields: group, axis, register, epistemic status, dependency edges, gates) | | |
| | `provenance/corpus_provenance.jsonl` | one row per document: writer model and batch, recast spec, source paper id, URL, dataset and license, verifier model with its relevance/polarity judgment, retrieval score | | |
| | `provenance/MANIFEST.json` | corpus-level summary and checks (writer models, verifier, group/relevance/polarity counts, source shares) | | |
| | `provenance/UNRESOLVED.jsonl` | documents whose source could not be rejoined (empty) | | |
| | `RECAST_SPEC.md`, `SCHEMA.md`, `schema.py` | the recast specification, schema and validator used to build the corpus | | |
| | `chunk_corpus.py` | the chunking used to make training records | | |
| ## Counts and known issues | |
| - The manifest lists 934 provenance rows for 932 distinct document ids (two rows are duplicates). | |
| `recast_documents.jsonl` contains 931 documents; the text of one manifest document | |
| (`recast_90512270`) could not be recovered from the recast outputs. | |
| - Groups: H→B 353, option value 265, B→A documented 313. Mean length about 470 words. | |
| - Source concentration is a known, failing check: peS2o supplied 71.0% of the documents against a | |
| 35% design cap, because a configured PubMed source could not be read and contributed nothing. | |
| - The training mixture used in the paper adds replay strata from | |
| [`Bioaligned/bioaligned22M`](https://huggingface.co/datasets/Bioaligned/bioaligned22M) and | |
| oversamples the three-axis documents; see the paper for the recipe. | |
| ## License and attribution | |
| The rewritten text is released under CC BY 4.0. Source papers are attributed per document in the | |
| provenance file (peS2o: ODC-BY 1.0; arXiv abstracts: per paper). Please cite the paper and the | |
| source papers you rely on. | |