Safetensors
English
German
speculative-decoding
eagle3
draft-model
kolibri
josch15366's picture
Kolibri-1 EAGLE-3 drafter, run 11 (72.9M training tokens on top of runs 1-10)
6430e5d verified
Raw History Blame Contribute Delete
382 Bytes
{
"architecture": "kolibri1-eagle3-drafter",
"target": "Aleph-Alpha/Kolibri-1",
"drafter": {
"hidden": 2560,
"heads": 20,
"head_dim": 128,
"ffn": 4096,
"eps": 1e-06,
"theta": 10000.0,
"layers": 1,
"taps": 3
},
"taps": [
44,
47,
49
],
"draft_vocab": "draft_vocab.json",
"draft_vocab_size": 32768,
"training_tokens": 72859196,
"training_steps": 48245
}