gary-neuron-chat / config.json
gary23w's picture
v4.2: L=8 cortex (ppl 11.2, ctx 192), conversational wrapper (yes/no, corrections, coreference, intros), bundled gary-neuron math
cf77b0f verified
Raw
History Blame Contribute Delete
843 Bytes
{
"model_type": "gary-neuron-chat",
"cortex": "gary-4-petite (4L/4H/96d/ctx128, vocab2048) fine-tuned on SODA dialogue",
"plastic_memory": "softmax-attention Hebbian read over conversation; ~6k params (Wq 96x64 + bN,rdec,g)",
"timescales": {
"fast": "within-conversation memory (no weight change)",
"slow": "sleep consolidation w/ experience replay"
},
"cortex_params": 656448,
"memory_params": 6147,
"results": {
"fact_recall_memory_on": 1.0,
"fact_recall_memory_off": 0.0,
"sleep_session_loss": [
5.73,
4.1
],
"sleep_base_ppl": [
29.4,
30.7
]
},
"dependencies": "numpy, tokenizers",
"n_layer": 8,
"n_params": 1109952,
"arch": "gpt-numpy E96 H4 L8 BLK192 + plastic hippocampus + episodic store + yes/no + corrections + gary-neuron parietal",
"val_loss": 2.4152,
"val_ppl": 11.2,
"block_size": 192
}