POG-E4B-v1 / v21 /adapter_config.json
LordAce9's picture
v2 program: v2.1 multi-dataset variant, GGUF layer-trim early-exit (bit-exact), server-mode extractor, 5-task BEIR sweep, binary/int8 compression study, full ablation writeup
d27cbf8 verified
Raw
History Blame Contribute Delete
300 Bytes
{
"d_model": 2560,
"n_layers": 3,
"n_slots": 6,
"d_attn": 512,
"d_gate": 256,
"d_ff": 3072,
"d_out": 2560,
"matryoshka_dims": [
256,
512,
1024,
2560
],
"backbone": "unsloth/gemma-4-E4B-it-qat-GGUF (UD-Q4_K_XL)",
"intercept_layers": [
28,
32,
36
]
}