Multimodal sweep results

Generated by the parallel training orchestrator.

Verdict

qwen2_5-vl-3b is the strongest variant with best_loss=0.4113 after 200 steps.

Comparison

variantstatusattemptsinterventionsstepsbest_lossfinal_lossbest_accfinal_acc
qwen2_5-vl-3bdone102000.41130.52550.91200.8866
paligemma2-3bdone102000.41630.41630.91420.8829
smolvlm2-2_2bdone102000.46140.55770.90520.8951
qwen2-vl-2bdone102000.46790.54520.90820.8709
moondream2done102000.46880.55330.90590.8751
tinyllava-1bdone102000.46920.52310.90820.8873
florence2-largedone102000.52830.52830.89930.8716
smolvlm-500mdone102000.53460.62890.89510.8854
internvl2-1bdone102000.53740.56840.89190.8877
llava-onevision-0_5bdone102000.54140.59070.89270.8704
smolvlm-256mdone102000.61800.64580.88270.8827
smolvlm2-256mdone102000.62300.64900.88100.8617
nanovlm-222mdone102000.72650.75530.85890.8537

Loss

loss curves

Accuracy

accuracy curves