AIM-LEP: Approximate Isometry Mapping with Language Extension Pipeline
LoRA adapters from AIM-LEP experiments on Qwen3-8B and Qwen3-0.6B. AIM initialization recovers 92% of original model performance vs FOCUS's 47%.
Variants
| Adapter | Model | Description |
|---|---|---|
| aim_8b | Qwen3-8B | AIM + LEP at 8B scale |
| aim_06b | Qwen3-0.6B | AIM + LEP at 0.6B scale |
Usage
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support