Qwen2.5-32B-Instruct SFT Faces Temporal Drift

This is the right instinct, but the real bottleneck is dataset design, not “more AI.” If the preprocessing and sampling are sloppy, the model will just learn old consulting noise with expensive weights.

• –Recency should be encoded in data selection and sampling, not only in a prompt tag; if 2026 matters more than 2008, the training mix needs to reflect that
• –Break long decks into phase-level supervision units instead of one giant mega-example so the model learns reusable strategy logic, not mushy summaries
• –Org charts and flowcharts need structure-preserving export, ideally as hierarchy-aware Markdown or JSON, not flattened OCR text that destroys parent-child relationships
• –The on-prem constraint makes extraction quality critical; every bad OCR page or duplicate “Final_Final_v2” deck becomes training debt
• –Qwen2.5-32B-Instruct plus LLaMA-Factory is a sensible stack, but the real lift will come from curation, deduping, and hard filtering of low-signal decks