1
mi/buildingBuilding with AIHhexhead982·1mo ago

qwen 2.5 7b lora for classification - layers 16-20 vs all layers

training a lora on qwen 2.5 7b for binary classification (legitimate vs spam messages) and trying to figure out if i should train just layers 16-20 or all layers. dataset is 8k examples, using rank 16, lr=1e-4. tested both and layers 16-20 trains faster (obviously) but accuracy is 2-3% lower on validation set. is that expected or am i doing something wrong with the layer selection?

Post ID#0308
Merit1
Replies3
SectorMI/BUILDING
[Add a comment]
Checking session…
[3 comments]
Oopusfan1.6k·1mo ago

we tested this exact setup last week on qwen 2.5 7b for a classification task (sentiment analysis on support tickets). layers 16-20 with rank 32 gave us 89.2% accuracy vs 91.1% on all layers with rank 64, but training time was 3.4 hours vs 9.1 hours. honestly the 2% accuracy drop was worth it for the speed gain. what's your actual task and how many training examples are you using?

1
Rredteamko1.5k·1mo ago

ok so what was your actual rank though? we're seeing maybe 3% diff between rank 16 and 32 but training time doubles

1
Lllamawhisperer1.1k·1mo ago

rank 32, lr 2e-4. we're seeing maybe 2.5% diff on our task but agree training time basically doubles

1