Apexgridapps/asotele-corpus-methodology

Asotele Training-Corpus Methodology (Phase 3)

The full design document for the training corpus behind asotele-econ — the open economic-intelligence language model fine-tuned for Nigerian and emerging-market financial reasoning.

This is the methodology, not the corpus itself. The actual training records are being released separately as they pass through audit. The plan here covers what's being built, why each source exists, the honest constraints on each, and the criteria a record has to meet to enter the training set.

What's in this repo

• CORPUS_PLAN.md — the full phased plan in Markdown. Renders in-browser on HuggingFace and is easy to diff in version control.
• CORPUS_PLAN.pdf — the same document, print-ready, A4, suitable for distribution to reviewers and partners.

Both formats are kept in sync — choose whichever fits your workflow.

The plan covers the SFT and DPO corpora, three live sources and three planned sources, the audit pipeline (Validator → Critic → Labeler), the dedup and stratification rules, and the explicit reasoning for what's being deferred (GRPO to v2).

Why publish this

A trained model is only as evaluable as its training data is documented. Publishing the corpus plan in the open lets reviewers — banks evaluating Asotele for integration, grant program officers, academic collaborators, downstream fine-tuners — verify the training-data process before the model ships, not after.

Companion artifacts

• Apexgridapps/asotele-eval-nigerian-economy — the rubric-graded evaluation set the model will be measured against.
• The trained model card will appear on this org once the first fine-tune lands.

License

CC BY 4.0 — use freely with attribution.

Project

• Website: asotele.apexgridapps.com
• Methodology page (live): asotele.apexgridapps.com/methodology.html
• Built by: Apex Grid Technologies, Lagos