What does fine-tuning change?

One lineage, several tracks.

Text generation, OCR/vision and classification need separate task protocols. Format conversions and quantized copies should be grouped with their source checkpoint; changed architectures belong in a separate comparison. A declared base-model tag is evidence of lineage, not proof of identical architecture or training.

Compare gains, not just rank.

Report each benchmark score and its difference from the base control, paired uncertainty, regressions, throughput and GPU-slot time. With almost identical model sizes, score versus latency and per-task gain charts are more useful than a crowded parameter-size scatter plot. The first measurement is all 2,376 ARC-Easy test questions, raw summed answer likelihood, no chat template or images. This measures text capability of OCR/vision specialists, not their OCR quality. Charts will appear after compatible results exist.

Keep the evaluation controlled.

Pin model revisions and item IDs. Use a common completion protocol for base capability; evaluate chat-native behaviour separately with explicit templates, thinking settings and equal token budgets. Count reasoning tokens inside generation budgets. Full cheap tests can run directly; adaptive estimates for expensive tests need calibration and a distinct protocol, followed by fixed-set confirmation.

How it fits the ontology

A cohort groups pinned System versions by declared lineage. Tasks and Benchmarks define what is measured; Result links a System to a fixed Protocol. A card claim is reported evidence, not a CodeSOTA measurement.