Qwen3.5-2B Early Exit Adapters
Adapters trained to map intermediate hidden states to final-model-like next-token distributions.
Base model: Qwen/Qwen3.5-2B
Layers: [4, 8, 12, 16, 18, 20]
Sequence length: 128
Training steps: 500
Loss: KL distillation to final logits
