BoBERT 路 v16-2026-11-12
Hit-object sequence encoders and retrieval indices for osu! modes. Each mode has its own tokenizer, encoder, layer-centering statistics and collection-trained linear adapter, all in model.safetensors.
| Mode | Maps | Dimensions | Layers |
|---|---|---|---|
| std | 745,963 | 512 | 9 |
| taiko | 96,102 | 384 | 9 |
| catch | 24,840 | 256 | 9 |
| mania | 315,286 | 384 | 9 |
<mode>/embeddings/ holds append-only bfloat16 shards; a later shard replaces earlier rows of the same beatmap. Each shard records the checksum of the model that embedded it. densities.parquet holds the retrieval density of every indexed map, and search.bin the client search catalog. Metadata catalogs are in data/.
Export workspace: 16c7703141140bdca72faff030d51d7ae81f9e4d.
From a source checkout:
import torch
from core.model import BobertEncoder
model = BobertEncoder.from_pretrained(
"runs/current/mania/model.safetensors", torch.device("cpu")
)
Inference Providers NEW
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support