ImmuneCoT-Qwen3-4B-ablation-poe-no-base (RQ3 ablation)
Ablation variant: poe_no_base (naive fusion, no Base-adjustment).
The full ImmuneCoT method fuses the two safety branches with a Base-adjusted
product-of-experts Q_imm(v) ∝ q̃_rec(v)·q̃_res(v)/q_B(v). This checkpoint uses
the naive fusion Q_no-base(v) ∝ q̃_rec(v)·q̃_res(v) — the same branch weights
(λ_rec=0.5, λ_res=0.7) but no division by the base distribution — isolating
whether the gains come from combining Recognition+Response at all, or
specifically from the Base-adjusted PoE term.
- Base model: Qwen3-4B (frozen teacher, full-parameter on-policy distillation)
- Data: iasd_data_decontam_v1.json (1571 samples), 2 epochs (198 steps)
- Objective: forward KL + per-token clip β=1.0 + top-k 256
- Fusion:
poe_no_base(P_immune = log Pd' + log Pr', no − log P_Ref)
Intended use: research reproducibility for the ImmuneCoT paper's RQ3 ablation.