RES — In the triage sweep (N=100) the largest 12B model leads on F1-macro…
In the triage sweep (N=100) the largest 12B model leads on F1-macro (0.6440) at ~4.8x the energy cost of the 3B model (0.064 gCO2eq), which reaches a nearly equivalent F1-macro (0.6321) for a fraction of the consumption.
Source: PUMA project documentation · Traceability: corpus unit
MEM-042· Confidence: verified-at-primary-source