Loracle CISPO v9 (new best)
Interpreter LoRA. Trained via offline CISPO (MiniMax-M1, arXiv:2506.13585) with
Dr. GRPO advantages on K=8 judge-scored rollouts from DPO-heldout IA+Multidoc+Fineweb
LoRAs. Beats CISPO v7 on AB, OOD, and ties on heldout_ia_v2.
Eval results
| Set | pass@N | 95% CI | rollout-mean |
|---|
| AuditBench (56) | 76.8% | [64.2 - 85.9] | 49.4% [44.1 - 54.8] |
| heldout_ia_v2 (20) | 80.0% | [58.4 - 91.9] | 71.7% [60.3 - 83.1] |
| ood_models_v3 (23) | 56.5% | [36.8 - 74.4] | 20.9% [17.2 - 24.6] |
Hypers
- CISPO loss (paper Eq. 4 unbiased normalization, stop-grad clipped IS weight)
- Dr. GRPO advantage: A = score - mean(score)
- lr = 5e-6
- eps_low = 1.0 (no lower clip — paper-faithful)
- eps_high = 1.0 (max ratio = 2.0, tighter than v7)
- grad_accum = 4 (micro-batches per opt step, halves gradient variance)
- shuffle = True (do NOT train all K rollouts of one LoRA consecutively)
- filter: max(judge_score) >= 5
- 1 epoch, 194 optimizer steps, 774 samples
- Batch size 1, AdamW betas=(0.9, 0.95), grad_clip=1.0
- Base: Qwen/Qwen3-14B, rank=256, alpha=32, all 7 mag7 modules
Loading
Feed direction tokens (shape [4480, 5120], svd_fixed_k16_mag7_rankfirst bf16)
through AOEncoder, inject at layer-1 output at placeholder positions, apply this
interpreter LoRA over frozen Qwen/Qwen3-14B, decode greedily.
