Quant matrix — masked val loss (gemma-4-31B base vs LoRAs)
Configs: base = no adapter; ensemble = all-speaker r=64 LoRA; others = dedicated per-character r=32 LoRAs. Each column is one 1-GPU job evaluating all 11 val sets (token-weighted mean masked NLL; lower is better). 2026-06-10.
| val set \ config | base | ensemble | spongebob | patrick | squidward | mr_krabs | plankton | sandy | mrs_puff | karen | gary | pearl |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| spongebob (n=1434) | 2.4966 | 1.8824 | 1.9136 | 1.9584 | 1.9599 | 1.9651 | 1.9601 | 1.9783 | 1.9920 | 1.9889 | 2.0040 | 1.9676 |
| patrick (n=881) | 2.6757 | 1.9520 | 2.0256 | 1.9917 | 2.0513 | 2.0569 | 2.0556 | 2.0744 | 2.0880 | 2.0822 | 2.1110 | 2.0538 |
| squidward (n=444) | 2.5061 | 1.9352 | 2.0122 | 2.0262 | 1.9790 | 2.0208 | 2.0136 | 2.0388 | 2.0484 | 2.0419 | 2.0688 | 2.0150 |
| mr_krabs (n=633) | 2.4928 | 1.9953 | 2.0992 | 2.1096 | 2.1075 | 2.0402 | 2.1025 | 2.1191 | 2.1442 | 2.1129 | 2.1208 | 2.0968 |
| plankton (n=581) | 2.5243 | 2.0330 | 2.1063 | 2.1252 | 2.1226 | 2.1239 | 2.0835 | 2.1313 | 2.1540 | 2.1393 | 2.1564 | 2.1140 |
| sandy (n=273) | 2.4949 | 1.9401 | 2.0092 | 2.0475 | 2.0569 | 2.0260 | 2.0431 | 2.0257 | 2.0604 | 2.0624 | 2.0560 | 2.0210 |
| mrs_puff (n=32) | 2.7816 | 1.9149 | 1.9684 | 2.0248 | 2.0255 | 2.0014 | 2.0366 | 2.0308 | 2.0280 | 2.0514 | 2.0945 | 2.0583 |
| karen (n=105) | 2.5326 | 1.9224 | 1.9952 | 2.0071 | 2.0124 | 1.9898 | 1.9900 | 2.0222 | 2.0180 | 2.0078 | 2.0320 | 1.9830 |
| gary (n=31) | 2.6714 | 1.3705 | 1.4656 | 1.4371 | 1.5072 | 1.4629 | 1.4397 | 1.4014 | 1.4782 | 1.4638 | 1.4448 | 1.4518 |
| pearl (n=24) | 2.4184 | 1.9298 | 1.9822 | 1.9696 | 1.9854 | 2.0194 | 2.0207 | 1.9778 | 2.0223 | 2.0143 | 1.9775 | 1.9404 |
| _ensemble (n=5392) | 2.5623 | 1.9588 | 2.0299 | 2.0438 | 2.0515 | 2.0456 | 2.0476 | 2.0674 | 2.0850 | 2.0723 | 2.0901 | 2.0502 |
| mean Δ vs base (11 sets) | +0.0000 | -0.6657 | -0.5954 | -0.5832 | -0.5725 | -0.5822 | -0.5785 | -0.5718 | -0.5489 | -0.5563 | -0.5455 | -0.5823 |
Ensemble vs dedicated, on each character's own val set
| val set | base | dedicated | ensemble | ens Δ vs base | ens Δ vs dedicated |
|---|---|---|---|---|---|
| spongebob | 2.4966 | 1.9136 | 1.8824 | -0.6142 | -0.0312 |
| patrick | 2.6757 | 1.9917 | 1.9520 | -0.7237 | -0.0397 |
| squidward | 2.5061 | 1.9790 | 1.9352 | -0.5709 | -0.0438 |
| mr_krabs | 2.4928 | 2.0402 | 1.9953 | -0.4975 | -0.0449 |
| plankton | 2.5243 | 2.0835 | 2.0330 | -0.4913 | -0.0505 |
| sandy | 2.4949 | 2.0257 | 1.9401 | -0.5548 | -0.0856 |
| mrs_puff | 2.7816 | 2.0280 | 1.9149 | -0.8667 | -0.1131 |
| karen | 2.5326 | 2.0078 | 1.9224 | -0.6102 | -0.0854 |
| gary | 2.6714 | 1.4448 | 1.3705 | -1.3009 | -0.0743 |
| pearl | 2.4184 | 1.9404 | 1.9298 | -0.4886 | -0.0106 |
| _ensemble (own set) | 2.5623 | — | 1.9588 | -0.6035 | — |
Sanity checks
- ensemble on squidward: 1.9352 vs known 1.935 -> OK
- ensemble on spongebob: 1.8824 vs known 1.882 -> OK
- degenerate cells (loss >5 or <0.5): none