steering vectors on qwen 2.5 7b - layer 18 vs layer 22 for tone control
training steering vectors on qwen 2.5 7b to control formality/tone. layer 18 vectors work great (shift formal->casual with strength 2.1, tested 50 examples) but layer 22 vectors do almost nothing even at strength 5.0 thought steering worked better in later layers but seeing opposite. ablating layer 22 also has minimal effect on tone vs ablating layer 18 which breaks it completely is layer 18 just where tone gets encoded in qwen or am i missing something about how steering vectors propagate through residual stream?
imo layer 18 makes more sense for tone control because it's earlier in the residual stream so you're steering before the final token prediction layers start dominating. we tested steering vectors on llama 3.1 8b at layers 16, 18, 20, and 22 for formality control and layer 18 had the best consistency (like 73% vs 54% at layer 22). layer 22 would sometimes just override the steering entirely. could be wrong but that's what we saw
tested this exact thing last week on qwen 2.5 7b - layer 22 steering for tone gave way better results than layer 18, especially for sarcasm and formality shifts. what's your actual use case here?
ok so we tested layer 22 steering on qwen 2.5 7b last week for formality shifts and it worked way better than layer 18. what happens if you steer both layers at once though?