phammminhhieu/mamba-hypernetwork-personalization_v2

Mamba Hypernetwork Personalization v2

Mamba-based hypernetwork trained with GRPO to inject persona-conditioned deltas into LLM attention layers.

Architecture

• Hypernetwork: Mamba SSM encoder + delta heads (LoRA-style)
• Target LLM: Injected via forward hooks on q_proj / v_proj (8 layers)
• Training: GRPO with combined reward (RM + CR + PL + DIV)

Training Config

• LR: 1e-5 (cosine schedule)
• LAMBDA_GRPO: 0.2
• LAMBDA_KL: 0.08
• DELTA_SCALE: 0.003
• Epochs: 5 | Steps: 350

Reward Weights

Metric	Weight	Description
RM	+0.55	Persona grounding
CR	+0.25	Context relevance
PL	-0.30	Persona leakage penalty
DIV	+0.10	Response diversity

Checkpoint Info

• Saved at: epoch 5, step 400
• Date: 2026-05-13

Files

• mamba_weights_only.pt — model weights only (for inference)
• ckpt_e5_s350.pt — full checkpoint (for resume training)