g4bal_e — RAI safety classifier (Atlassian responsible-ai campaign)

• Base: google/gemma-4-E2B-it
• Method: full-SFT (final merged model; optimizer states not saved by the trainer)
• Arm / iteration: g4bal_e / g4bal
• Contents: model weights (root) · eval/ = raw inference outputs on the eval sets (v14, hello_v2_filtered, v200, robustnessv2) under policy prompts v3_1b / v3_1c · artifacts/ = training log
• Eval-inference files uploaded: 10