Gemma 4 E2B-it Pi-Mono SFT — youtube-livestream-1
Merged full weights selected from a three-run LoRA sweep by minimum held-out
session-level evaluation loss.
- Base:
google/gemma-4-E2B-it
- Selected adapter:
dotemunk/gemma-4-E2B-it-pi-mono-lr2e4-r16-a32-youtube-livestream-1
- Held-out eval loss:
0.5826800465583801
- Split: source session, zero train/eval session overlap
- Trackio: local Trackio project
youtube-livestream-1
The final Inspect AI score table is added after HumanEval and MBPP complete.
