Supermix v101 (v101_scf_cognitive / Cognitive Fabric)

Standalone release for v101_scf_cognitive.pt and the v101 Cognitive Fabric runtime from kai9987kai/Supermix (main commit 1404afbd), also cross-packaged inside Kai9987kai/supermix-v99-cognitive-fabric.

What v101 Adds

  1. Continuation Neural Checkpoint (weights/v101_scf_cognitive.pt):
    • Warm-started from Kai9987kai/supermix-v100 (v100_scf_balanced.selected.pt) on the 10,100-row datasets/v101/v101_cognitive_mix.jsonl (+45 in-vocab growth IDs -> 9,460 vocabulary; 36,614,279 total / 9,147,383 active params).
    • Achieves lowest lineage best_dev_loss = 0.01745 (best_step = 32, passed = True; tier1_seen_response 0.00497 / ppl 1.005 on 32 pairs; tier3_unseen_sentence 0.051251 / ppl 1.0526 on 364 pairs).
  2. Zero-eval Extended Symbolic Mathematics (source/nexus_solver.py, source/scf/verification/):
    • Implicit-coefficient 2x2 linear systems & quadratics, Newton finite-difference (degree 1..3) / geometric / Fibonacci sequence discovery, polynomial derivatives & definite integrals, gcd/lcm/median/variance/std_dev, safe AST evaluator (_eval_safe_ast_expr), and algebraic equivalence checking with verification certificates & derivation steps.
  3. Structured Problem Solving, Trace Auditing & Ambiguity Probing (source/scf/reasoning/, source/scf/collective/, source/scf/routing/):
    • Step decomposition with ambiguity probes, mid-trace arithmetic auditing (_audit_trace_arithmetic), and prose problem-solving submode routing (SCFRouter.assess_task).
  4. MAP-Elites + NoveltySearch Ideation (source/nexus_ideation.py):
    • Topic-grounded SCAMPER/TRIZ/Analogy/Contrarian ideation selected across min-max normalized 2D MAP-Elites behavioral cells + Jaccard novelty distance with falsification hypotheses.
  5. Memory-Aware Multi-Turn Continuity & Upgraded Live UI (source/scf/memory/, source/scf/core/, source/supermix_chat_server.py):
    • Importance-scored conversational constraint capture (Memory alignment:) with contradiction detection, assistant_synthesis UI cards (VERIFIED / ADVISORY), and live telemetry.

Honest Benchmarks & Rule-9 Status (Windows CPU-only)

All evaluations measured on Windows, PyTorch 2.11.0+cpu, cuda_available=False, max_new_tokens=96:

Evaluation Suite Result Notes
Standalone Neural seed=82 (41 tasks) 35 / 41 (0.8537) results/eval_v101_seed82.json
Standalone Neural seed=83 (82 tasks) 72 / 82 (0.8780) results/eval_v101_n82_seed83.json
Standalone Neural seed=84 (82 tasks) 71 / 82 (0.8659) results/eval_v101_n82_seed84.json (178 / 205 = 86.8% combined)
SCF Hybrid (python -m scf bench --seed 82 --neural) 41 / 41 (1.0000) Deterministic verification + connectome atlas

Per Rule 9 promotion gating in kai9987kai/Supermix, v96_physio_neurogenesis.selected.pt (180 / 205 = 87.8%) remains #1 Promoted Default and v100_scf_balanced.selected.pt (179 / 205 = 87.3%) #2; v101_scf_cognitive.pt is released as Experimental (#3 Standalone Accuracy, Lowest Dev Loss).

Quick Start

pip install torch numpy flask huggingface_hub
huggingface-cli download Kai9987kai/supermix-v101 --local-dir .

# 1. Run standalone v101 inference + SCF verification + ideation demo
python example_usage.py

# 2. Run Cognitive Fabric CLI
PYTHONPATH=source python -m scf status
PYTHONPATH=source python -m scf bench --seed 82 --neural

# 3. Launch live chat web UI with v101_scf_cognitive + SCF Synthesis cards
python source/supermix_chat_server.py --model v101_scf_cognitive=weights/v101_scf_cognitive.pt --port 8780
Downloads last month
43
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Kai9987kai/supermix-v101