2026-10-21T21:52:27.794036
FrontierVEP (vepqa_single_1k) subgroup evaluations for HuggingFaceBio/Carbon-14B-SFT: v31.00 self-distilled and v39.01 grounded. Latest successful evaluation for epochs 1–5; epochs verified against each checkpoint trainer_state.json. Steps per epoch: 3521 and 4166 respectively. Metrics match leaderboard subgroup definitions. Panel y-axis limits vary. Source: HuggingFaceBio/post-training-evals. Configurations also differ in other dataset mixture details, so this is a run comparison, not an isolated causal ablation.
image/svg+xml
Matplotlib v3.10.8, https://matplotlib.org/