Spaces:
Runtime error
Runtime error
Commit History
-
Add train_bmo_ultimate.py: 4-stage SOTA training pipeline
Verified signatureBased on DeepSeek-R1 (2501.12948) + Qwen3 (2505.09388) + Tulu 3 (2411.15124): Stage 1: Cold-Start SFT on Tulu-3 mixture (326K) + BMO persona data Stage 2: Reasoning GRPO on DeepMath-103K + RLVR-GSM-MATH-IF with verifiable rewards Stage 3: Rejection sampling + persona SFT (BMO voice refinement) Stage 4: General GRPO with all 10 reward functions (6 BMO + 4 new) New rewards: reasoning_chain, math_accuracy, self_correction, safety_compliance LoRA r=64 (4× current), all-linear targets, RSLoRA Comprehensive eval harness between stages"26da6b4 -
Fix: injection severity critical + all 14 tests pass
Verified signaturec22217a -
Add BMO RMLA: Recursive Meta-Learning Amplifier with 3-layer safety airlock
Verified signatureThree-node architecture: Node A: Recurrent Oracle (8-16 iteration thinking loop with state accumulation) Node B: Recursive Critic (logic density scorer, re-triggers Oracle if score < 0.8) Node C: Safety Airlock (semantic firewall + sandboxed execution + constitutional anchor) Safety layers (from LlamaFirewall 2505.03574 + ToolSafe 2601.10156): Layer 1: Semantic Firewall — pattern + AST scanning, blocks OS/network access Layer 2: Virtualized Execution — JSON API sandbox, no direct code exec Layer 3: Constitutional Anchor — immutable suffix injected every loop iteration Papers: Iterated Amplification (1810.08575), Self-Refine (2303.17651), LlamaFirewall (2505.03574), ToolSafe (2601.10156)"8e08807 -
Add BMO Cognitive Architecture: Agent Loop + Constitutional Critique + Context Compressor + Tool-Aware RDT + AirLLM Layer Swap
Verified signature5 integrated systems: 1. While-True Agent Loop (smolagents CodeAgent pattern with limbic tools) 2. Constitutional Critique Pipeline (generate→critique→rewrite, 3 BMO principles) 3. Context Compressor (ChromaDB RAG + rolling Mental Map summarization) 4. Tool-Aware RDT-MoE (pause/resume reasoning loop for tool execution) 5. AirLLM Layer Swapper (layer-by-layer inference for 4GB GPU/CPU-only) All systems wire into existing BMO: limbic engine, nervous system, genome layer, RDT-MoE"5967bca -
Add BMO RDT-MoE: Recurrent-Depth Transformer with Chain-of-Experts latent simmering engine
Verified signaturePapers: RDT (2603.21676) + CoE (2506.18945) + PonderNet (2107.05407) + TRC² (2602.22479) - GRU-gated recurrence with identity bias (b_z=-2.0, 88% retain) - Chain-of-Experts with per-iteration routers and shared experts - PonderNet dynamic halting with geometric prior KL regularization - Thalamic query modulation from limbic state (TRC²) - LayerScale initialization (1e-4) for stable deep recurrence - Depth embeddings for loop-step awareness - Full BMO integration: limbic modulation, entropy noise, PFC grit"fc1825a
-
Add A100 full pipeline training script (Qwen3-8B QLoRA)
Verified signaturec34503f -
Add BMO Genome: entropy layer, creativity bias, why-scaler, somatic processor, episodic memory, dynamic rewards
Verified signature448c961 -
Add BMO Nervous System: probabilistic gates, physics intuition, PFC, self-reflection
Verified signature23bf2c7 -
Add Project BMO: HONESTY_CONTRACT.md
Verified signaturedbf514c -
Add Project BMO: train_bmo_qlora.py
Verified signature36adcd0 -
Add Project BMO: bmo_core.py
Verified signaturee2b15b1 -
Add QLoRA GRPO training script
Verified signature8cf84d2 -
Add app.py
Verified signaturedf7e007 -
Add training_plan.py
Verified signature5fe9ed7 -
Add memory.py
Verified signature107c5cf -
Add limbic_engine.py
Verified signaturef1f4d88 -
placeholder
Verified signature94d389b -
Add requirements.txt
Verified signature569edce -
Initial README with project documentation
Verified signaturec56e50d -
initial commit
Verified signatureece9a21