AdaAnchor-KV — Qwen2.5-1.5B (Lightweight Weights)
This repository contains the trained lightweight parameters for AdaAnchor-KV + Learned Halting.
Backbone (frozen during training):
Qwen/Qwen2.5-1.5B-Instruct
What's inside
adaanchor_kv_weights.pt:
anchors: learnable anchor vectors (m=8)
proj: projection layer for anchor refinement
ln: layer norm for refinement update
halt: learned halting head (stop probability)
config.json: training + method hyperparameters
Training summary
- Kmax: 6
- beta: 0.5
- lambda_cost: 0.02
- max_q_len: 192
- max_a_len: 32
Note: These are NOT full model weights. Load the backbone separately with Transformers,
then load these lightweight weights to run AdaAnchor-KV refinement with KV caching.