Supermix v100 (v100_scf_balanced)

Standalone release for v100_scf_balanced.selected.pt from kai9987kai/Supermix (main branch), also cross-packaged inside Kai9987kai/supermix-v99-cognitive-fabric.

Architecture & Training Summary

  • Checkpoint: weights/v100_scf_balanced.selected.pt (153,161,606 bytes)
  • Parameters: 36,599,879 total / 9,132,983 active per token (9,415 vocabulary, 6 transformer layers)
  • Connectome & Physiology Extensions: Adds default-off Carandini-Heeger E/I divisive normalization (cns_divisive_norm) over Dale inhibitory pools alongside cns_predictive_error, cns_hebbian_trace, and thinking_entropy_halt, warm-started from v96_physio_neurogenesis.selected.pt with frozen trunk scheduling (12 steps), no_norm_bias weight decay, MTP decay, and hard-subset rehearsal.
  • Training Corpus: datasets/v100/v100_precision_mix.jsonl (8,470 rows, sha256: dd7dca8d119763279973297781e72dda1cde09211f89e44ca6d48be1cbe522c7).
  • Generalisation & Ablations (results/generalisation_results.json):
    • best_step = 16, best_dev_loss = 0.023837, passed = True
    • tier1_seen_response (62 pairs): loss 0.01004 (ppl 1.0101)
    • tier3_unseen_sentence (761 pairs): loss 0.019318 (ppl 1.0195)
    • Connectome ablations (32 rows, ref 0.017653): gates_off +0.000238 nats, left_only +5.9e-05 nats, right_only +0.000246 nats

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_v100_selected_seed82.json
Standalone Neural seed=83 (82 tasks) 72 / 82 (0.8780) results/eval_v100_selected_n82_seed83.json
Standalone Neural seed=84 (82 tasks) 72 / 82 (0.8780) results/eval_v100_selected_n82_seed84.json (179 / 205 = 87.3% 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%, 73 / 82 = 0.8902 on seed=84) remains the Promoted Default baseline, while v100_scf_balanced is released as Experimental (#2 Standalone Neural Accuracy).

Quick Start

pip install torch numpy flask huggingface_hub
huggingface-cli download Kai9987kai/supermix-v100 --local-dir .

# 1. Run standalone v100 inference + SCF verification 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 v100_scf_balanced
python source/supermix_chat_server.py --model v100_scf_balanced=weights/v100_scf_balanced.selected.pt --port 8780
Downloads last month
45
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Kai9987kai/supermix-v100

Finetuned
(1)
this model
Finetunes
1 model