Release v15

README.md

@@ -6,43 +6,35 @@ tags:
 - similarity-search
 ---
 
-# BoBERT · v14.1
+# BoBERT · v15
 
-Dense osu!standard beatmap representations from `.osu` hit-object sequences.
-The encoder was pretrained on roughly 500,000 beatmaps; collection-derived adaptation
-refines the embeddings for recommendation and similarity search.
+Hit-object sequence encoders and retrieval indices for osu! modes. Each mode has its own tokenizer, encoder weights, centering statistics and optional collection-trained linear adapter.
 
-[Source and architecture](https://github.com/token03/bobert) · [Live recommender](https://bobert-web.pages.dev/)
+| Mode | Variant | Maps | Dimensions | Layers |
+|---|---|---:|---:|---:|
+| std | adapted | 498,962 | 384 | 9 |
+| taiko | adapted | 95,415 | 384 | 9 |
+| catch | adapted | 24,550 | 256 | 9 |
+| mania | adapted | 314,997 | 384 | 9 |
 
-## Release
+The release manifest selects one model/index pair per mode. Base and adapted files use the same mode directory; adapted files have an `.a` suffix. The index records its model's SHA256 checksum. Catalogs are shared in `data/`.
 
-- Embeddings: 498,962 beatmaps, 384 dimensions.
-- Encoder: 9 layers, 6 heads, 4,096 hit-object limit.
-- Index generated: 2026-09-14T15:25:19.001095+00:00.
-- Adapter: collection-trained linear projection.
-- Export workspace: `a65b9137198c8e64378960a881d66036fd4fbe19`.
-
-## Download
-
-From the source checkout:
+Export workspace: `f77d80d8442a535c3d7700f245a11022157dc89c`.
 
 ```sh
-uv run --no-default-groups --group serve fetch-run --repo token03/bobert --revision v14.1
+uv run fetch-run --repo token03/bobert --revision v15
 docker compose up --build api
 ```
 
-To inspect the model directly:
+From a source checkout:
 
 ```python
 import torch
 from core.model import BobertEncoder
+from core.artifacts import resolve_artifacts
 
-model, stats = BobertEncoder.from_pretrained("model.safetensors", torch.device("cpu"))
+artifacts = resolve_artifacts("runs/current", "mania")
+model = BobertEncoder.from_pretrained(artifacts.model, torch.device("cpu"))
 ```
 
-`model.safetensors` contains the encoder, adapter, normalization statistics, and model configuration.
-`embeddings.parquet` contains beatmap IDs, vectors, densities, and embedded pooling/retrieval metadata.
-The index records its model's SHA-256 checksum. Keep the two files together for online queries.
-`data/` contains the metadata and strain catalogs required by the API.
-
-See the source repository for the training pipeline, acknowledgements, limitations and evaluation tools.
+[Source, training and evaluation](https://github.com/token03/bobert)

model.safetensors → catch/embeddings.a.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:40aa11d62aa8d88ee987932847acf61d8a751d8b654b51f82018e4a699708d55
-size 64609608
+oid sha256:ec5303b3d5c843626fc4c01089dcf181e93efcf2aedc018a77e0956736a16a78
+size 10061295

embeddings.parquet → catch/model.a.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7f92bcd664f86500d5a32c5f2ccc69957ae35db792bb30064ea198deb9ead30e
-size 384237851
+oid sha256:3e96d5d2c4c1eb507f6a3c87d12778768abdef7a62102efb264903c883adbabb
+size 28900032

catch/training.a.yaml

@@ -0,0 +1,58 @@
+data:
+  mode: catch
+  features_dir: /home/jessiez/projects/bobert/data/features
+  strains_path: /home/jessiez/projects/bobert/data/strains.parquet
+  dataset_seed: 42
+  sample_size: null
+  max_seq_len: 4096
+  val_split: 0.1
+  min_sr: 2.0
+  max_sr: 14.0
+  dataloader:
+    num_workers: 2
+    pin_memory: true
+    persistent_workers: false
+    prefetch_factor: 2
+  validation_metadata_path: /home/jessiez/projects/bobert/runs/v15/catch/validation_groups.parquet
+  validation_group_column: validation_group_id
+model:
+  d_model: 256
+  n_heads: 4
+  n_layers: 9
+  dim_feedforward: 688
+  dropout: 0.1
+  local_attention_window: 128
+  local_attention_block_size: 256
+  global_attention_layers:
+  - 2
+  - 5
+  - 8
+  feature_token_dim: 16
+runtime:
+  compile_model: true
+  compile_mode: default
+  activation_checkpointing: false
+training:
+  seed: 42
+  trainer:
+    epochs: 30
+    batch_size: 64
+    max_tokens: 32768
+    grad_accum: 2
+    grad_clip: 1.0
+    precision: bf16-mixed
+    save_checkpoints: true
+  optimizer:
+    adamw_lr: 0.0002
+    adamw_wd: 0.05
+    adamw_betas:
+    - 0.9
+    - 0.95
+    muon_lr: 0.01
+    muon_wd: 0.01
+  scheduler:
+    warmup_ratio: 0.01
+    min_lr_ratio: 0.005
+  masking:
+    ratio: 0.3
+    mean_span_length: 2

data/beatmaps.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a5a3cfbe7916a63629e7411f71a8ec8c26535dcf9a8c56694262298d8680e15d
-size 80239403
+oid sha256:12f63cf72710b42acd8e48f28c589c7e2ee45758699750de9a4f9d7b1d3bf481
+size 214545102

data/strains.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:748f8e6bdab690dfc36604a4bfb78f8e0ddbcfd8faeb6c4ee9bab1b4d2b2cd1e
-size 40110268
+oid sha256:212b7c1587fac5d64ce806f7fac8c1d09577a14135e8ec1d67cfbcb6eb794216
+size 70206325

mania/embeddings.a.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6875387785899c2393d5bbfdca7d74cc82864e8a86a1ee941416f4807f8cb5d2
+size 189563261

mania/model.a.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3bc9148983d51c8273805d87a241ebbb039ae20c2096190e4f053c3104ae7b5
+size 64678760

mania/training.a.yaml

@@ -0,0 +1,58 @@
+data:
+  mode: mania
+  features_dir: /home/jessiez/projects/bobert/data/features
+  strains_path: /home/jessiez/projects/bobert/data/strains.parquet
+  dataset_seed: 42
+  sample_size: null
+  max_seq_len: 4096
+  val_split: 0.1
+  min_sr: 2.0
+  max_sr: 14.0
+  dataloader:
+    num_workers: 2
+    pin_memory: true
+    persistent_workers: false
+    prefetch_factor: 2
+  validation_metadata_path: /home/jessiez/projects/bobert/runs/v15/mania/validation_groups.parquet
+  validation_group_column: validation_group_id
+model:
+  d_model: 384
+  n_heads: 6
+  n_layers: 9
+  dim_feedforward: 1024
+  dropout: 0.1
+  local_attention_window: 128
+  local_attention_block_size: 256
+  global_attention_layers:
+  - 2
+  - 5
+  - 8
+  feature_token_dim: 16
+runtime:
+  compile_model: true
+  compile_mode: default
+  activation_checkpointing: false
+training:
+  seed: 42
+  trainer:
+    epochs: 12
+    batch_size: 64
+    max_tokens: 32768
+    grad_accum: 2
+    grad_clip: 1.0
+    precision: bf16-mixed
+    save_checkpoints: true
+  optimizer:
+    adamw_lr: 0.0002
+    adamw_wd: 0.05
+    adamw_betas:
+    - 0.9
+    - 0.95
+    muon_lr: 0.01
+    muon_wd: 0.01
+  scheduler:
+    warmup_ratio: 0.01
+    min_lr_ratio: 0.005
+  masking:
+    ratio: 0.3
+    mean_span_length: 2

release.json

@@ -0,0 +1,9 @@
+{
+  "format": 1,
+  "modes": {
+    "std": "adapted",
+    "taiko": "adapted",
+    "catch": "adapted",
+    "mania": "adapted"
+  }
+}

std/embeddings.a.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:974e354a7c20f059cbfd0bf8aa9282f49884403a191f0049571641fe5c754535
+size 316200943

std/model.a.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:511cb211efde9a13cb68fb7f6b28f3d19b0642f303ae90ade88d46205fb072d3
+size 64704648

taiko/embeddings.a.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e36ebb77fdbacecf70f82f57e220684d82e5c92b196273ef1489ef259ed4175
+size 57491033

taiko/model.a.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c324b5d5cfd69d69c32200c9a68b33c3d070637a63cb64fa901883b3356477aa
+size 64478384

taiko/training.a.yaml

@@ -0,0 +1,58 @@
+data:
+  mode: taiko
+  features_dir: /home/jessiez/projects/bobert/data/features
+  strains_path: /home/jessiez/projects/bobert/data/strains.parquet
+  dataset_seed: 42
+  sample_size: null
+  max_seq_len: 4096
+  val_split: 0.1
+  min_sr: 2.0
+  max_sr: 14.0
+  dataloader:
+    num_workers: 2
+    pin_memory: true
+    persistent_workers: false
+    prefetch_factor: 2
+  validation_metadata_path: /home/jessiez/projects/bobert/runs/v15/taiko/validation_groups.parquet
+  validation_group_column: validation_group_id
+model:
+  d_model: 384
+  n_heads: 6
+  n_layers: 9
+  dim_feedforward: 1024
+  dropout: 0.1
+  local_attention_window: 128
+  local_attention_block_size: 256
+  global_attention_layers:
+  - 2
+  - 5
+  - 8
+  feature_token_dim: 16
+runtime:
+  compile_model: true
+  compile_mode: default
+  activation_checkpointing: false
+training:
+  seed: 42
+  trainer:
+    epochs: 20
+    batch_size: 64
+    max_tokens: 32768
+    grad_accum: 2
+    grad_clip: 1.0
+    precision: bf16-mixed
+    save_checkpoints: true
+  optimizer:
+    adamw_lr: 0.0002
+    adamw_wd: 0.05
+    adamw_betas:
+    - 0.9
+    - 0.95
+    muon_lr: 0.01
+    muon_wd: 0.01
+  scheduler:
+    warmup_ratio: 0.01
+    min_lr_ratio: 0.005
+  masking:
+    ratio: 0.3
+    mean_span_length: 2