Files included
adapter_config.json & adapter_model.safetensors (158.3 MB, BF16): Standard PEFT LoRA targeting all decoder linear projections (self_attn.{q,k,v,o}_proj, linear_attn.{in_proj_qkv,in_proj_z,in_proj_b,in_proj_a,out_proj}, mlp.{gate,up,down}_proj).
- Shared-
A fused-projection layout: Within each layer, (q_proj, k_proj, v_proj), (in_proj_qkv, in_proj_z), (in_proj_b, in_proj_a), and (gate_proj, up_proj) share bitwise-identical lora_A.weight tensors from joint SVD factorization, allowing fused-projection engines (warp-34) to run a single rank-128 shrink per fused GEMM instead of 2r/3r.
- Frozen norms & head:
input_layernorm and post_attention_layernorm scale ratios ((1 + g_glide) / (1 + g_base)) are folded directly into the downstream lora_A columns so base-model RMSNorm weights remain untouched at runtime.
decision_config.json & readout.safetensors: 255-way CodeReadout head and noncausal_full_attention (quick-empty-think-medium-v1) metadata from fastino/GLiDE-0.8B for warp-34's per-adapter DecisionModelProfile.
JevBench Public Evaluation (231 cases: 48 Easy, 72 Original, 111 Hard)
Evaluated on kshitijthakkar/nirnaya-jevbench-public-eval-v1:
Table with columns: Configuration, Overall, Easy (48), Original (72), Hard (111), Mean NLL, Mean Brier| Configuration | Overall | Easy (48) | Original (72) | Hard (111) | Mean NLL | Mean Brier |
|---|
Qwen/Qwen3.5-0.8B + Qwen3.5-0.8B-GLiDE-LoRA (CodeReadout + noncausal_full_attention) | 187/231 (80.95%) | 48/48 | 70/72 | 69/111 | 0.4129 | 0.2382 |
Extraction method: Activation-Weighted Joint SVD (Joint-ASVD) with exact GemmaRMSNorm column-ratio folding.