Commands
/linguistic:findings
List structured findings (data gaps, ethics flags, fertility issues, eval contamination).
Syntax
/linguistic:findingsDescription
Surface structured findings accumulated by specialist skills during the session, grouped by severity:
| Severity | Examples |
|---|---|
| HIGH | License violation risk; sacred-text in training corpus; eval set contamination >10%; tokenizer fertility >5× baseline |
| MEDIUM | Significant register imbalance (e.g., Bible >50%); URIEL distance to chosen transfer source >0.6; missing ethics sign-off |
| LOW | Minor dedup gain available; stale benchmark version; non-canonical romanization |
Each finding includes: source skill, evidence, recommended action, and optional auto-fix toggle.
Example
/linguistic:findings
HIGH (1):
[linguistic-corpus] Bible-NLP Yoruba: liturgical register 62% of corpus
Evidence: register audit shows >30% threshold exceeded
Action: reduce Bible slice to ≤30%; supplement with web/news
MEDIUM (1):
[linguistic-eval] FLORES-200 detected in base model pretrain mix
Evidence: Llama-3 cutoff March 2024 includes FLORES
Action: report FLORES as lower bound; use NTREX-128 as primary
LOW (0): noneRelated Commands
/linguistic:review— full review including findings/linguistic:status— findings count in one-line summary
Was this page helpful?
Edit on GitHub
Last updated on