القائمة
Navigate mechanistic interpretability concepts, tools, and SAE architectures. Quick reference for SAE variants, feature steering methods, circuit discovery, and trust calibration.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Validate GRPO training configs against 36 cross-framework bug avoidance rules (19 MUST DO + 17 MUST NOT). Covers 7 frameworks (DeepSpeed, Megatron, vLLM, verl, SGLang, rLLM, MindIE), 7 pattern classes, and MoE NaN avoidance.
Debug and diagnose GRPO training issues across verl, Megatron, DeepSpeed, and RTX 4090. Symptom-to-bug mapping, NaN debugging, Muon clipping avoidance, and CUDA stream safety analysis.
Navigate SGLang source code. Quick reference for subsystem locations, key classes, and data flow.
Navigate model merging strategies and theory. Quick reference for Task Arithmetic, DARE, TIES merging methods, alpha selection, and RTX 4090 validation results.
Navigate RDMA, InfiniBand, RoCE v2, AI cluster networking topology, and NCCL transport internals. Quick reference for Verbs API, QP/CQ/SRQ/MR, fat-tree/rail-aligned topology, DCQCN congestion control, GPUDirect RDMA, and cluster configuration.
Navigate PyTorch torch.distributed, NCCL, and distributed training internals. Quick reference for ProcessGroupNCCL, DDP, FSDP, NCCL channels/proxy/protocol, and training configuration.
| name | interpretability-nav |
| description | Navigate mechanistic interpretability concepts, tools, and SAE architectures. Quick reference for SAE variants, feature steering methods, circuit discovery, and trust calibration. |
| triggers | ["interpretability","SAE","sparse autoencoder","mechanistic interpretability","feature steering","refusal circuit","circuit discovery","ACDC","attribution patching","causal scrubbing","ActAdd","activation steering","RepE","trust calibration","sycophancy","TopK SAE","Gated SAE"] |
Layer 1: SAE → decompose activations → features → interpretable units
Layer 2: Circuit Discovery → connect features → circuits → behavior mechanisms
Layer 3: Feature Steering → manipulate features → control behavior → applications
| Gen | Architecture | L0 Control | Dead Features | Absorption | Shrinkage | 2025 Rec |
|---|---|---|---|---|---|---|
| Standard | h=ReLU(W_enc*x+b) | Indirect(λ) | Yes | Yes | Yes | Baseline |
| Gated | gate=ReLU(W_gatex), mag=ReLU(W_magx), h=gate*mag | Better | Reduced | Reduced | No | Frontier |
| TopK | h=TopK(W_enc*x+b, K) | Exact=K | None | None | No | ✓ Best |
tools/sae_simulator.py| Method | Approach | Stability | Complexity | Production |
|---|---|---|---|---|
| ActAdd | s=a⁺-a⁻ contrastive difference | Low(brittle) | Lowest | Research |
| CAA | Average N contrastive pairs | Medium | Low | Research |
| RepE | PCA directions + multi-layer | Good | Medium | Possible |
| DiffSteer | Functional/task-specific | Good | Medium | Frontier |
| LoRA-Steer | Learned low-rank adapter | ✓Best | High | ✓ Production |
ACDC (auto prune edges) → Attribution Patching (gradient 100x faster) → Causal Scrubbing (strict verification)
Fast screening → Approximate refinement → Gold-standard confirmation
Harmful request → detection_features → intermediate_features → suppress_output → "I cannot help"
Dual-use: transparency vs security tension. Publishing refusal circuit details helps adversaries bypass safety.
tools/trust_calibration_simulator.pynotebook/fundamentals/mechanistic-interpretability-deep-dive.md — comprehensive theorynotebook/fundamentals/ai-safety-guardrails-production-deep-dive.md — Defense-in-Depth connectionnotebook/fundamentals/ai-economic-impact-human-ai-interaction-deep-dive.md — trust/sycophancy/anthropomorphism