Native training step-zero initialization for the 100M 20-prediction model. This exactly uses origin seed 11101994 and historical source commit fad735c90bd44b4bcfd6bdf9429edbb640d2ebf5, matching the start of W&B run 3g8jxx2k. The weights are untrained FP32 values produced before XLA/FSDP wrapping.

Model Details

Parameter	Value
Parameters	~0.22B
Layers	12
Hidden Size	768
Attention Heads	12
Vocab Size	8192
Prediction Steps	20

Usage

python

from transformers import AutoModel, AutoConfig

# Load with trust_remote_code for custom model
model = AutoModel.from_pretrained(
    "TuKoResearch/AuriStream100M_20Pred_BigAudioDataset_500k-randinit",
    trust_remote_code=True,
)

# Or load config first
config = AutoConfig.from_pretrained("TuKoResearch/AuriStream100M_20Pred_BigAudioDataset_500k-randinit", trust_remote_code=True)

Base Model Code

This checkpoint uses shared model code from TuKoResearch/AuriStream-base.

Tokenizer

This model uses cochlear tokens from WavCochCausalV8192.