Run log
9 runs, newest first.
Each row opens the full record of that run. Runs that failed or were abandoned stay in this list.
Table 1 · all runs
| Recorded | Kind | Run | Result |
|---|---|---|---|
| Repair eval | Qwen3.5-4B, untuned, 7 tasks | 6 of 21 trials passed | |
| Repair eval | Qwen3.5-4B, untuned, 10 tasks | 20 of 30 trials passed | |
| Lint baseline | Supabase Splinter, 17 tasks | 8 of 14 vulnerable tasks flagged | |
| LoRA smoke test | Qwen3.5-4B, sequence 2048, padded | 9.12 GiB peak, 142 s for 30 steps | |
| LoRA smoke test | Qwen3.5-4B, sequence 2048 | 8.81 GiB peak, 49 s for 30 steps | |
| LoRA smoke test | gemma-4-E4B-it, sequence 512 | 7.09 GiB peak, 231 s for 100 steps | |
| LoRA smoke test | Qwen3.5-4B, sequence 512 | 8.81 GiB peak, 159 s for 100 steps | |
| First load | gemma-4-E4B-it | 7.09 GiB peak, 28 s to load | |
| First load | Qwen3.5-4B | 8.72 GiB peak, 149 s to load |
Fig. 3 · peak VRAM against the card
Scaleeach bar runs from 0 to 12 GiB, the whole of one RTX 4070 Ti.