| OOM during training | SAE too wide for available VRAM | Reduce expansion factor, increase batch accumulation, or use Q4 model |
| OOM during steering | Model + SAE + KV cache exceeds VRAM | Use smaller SAE width or more aggressive quantization |
| >50% dead neurons | Sparsity too aggressive | Reduce l1_alpha/sparsity_coeff, enable sparsity warmup |
| Labels say "uncategorized" | LLM couldn't interpret the feature | Increase max_examples, try a larger LLM, check activation examples manually |
| Training loss spikes | Learning rate too high | Reduce by 2–5x, increase warmup steps |
| Training loss plateaus | Learning rate too low or not enough steps | Increase LR or total_steps |
| Labeling timeouts | Local model too slow for batch | Reduce batch_size to 1, increase api_timeout |
| Steering has no effect | Strength too low or wrong feature | Increase strength (try 20–50), verify feature has clear activation pattern |