File size: 5,912 Bytes
a4eaf45
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
---
language:
- en
pretty_name: "Beat Saber ranked maps — derived statistics & pattern priors"
license: other
tags:
- beatsaber
- beat-saber
- rhythm-game
- game-data
- json
- procedural-generation
---

# Dataset / model card (Beat Saber ranked maps)

This folder holds **dataset-style artifacts** produced by a **Python** training / aggregation pipeline over **3,000+ different Beat Saber song maps** (ranked community maps, parsed from many map archives). This README is the **Hugging Face dataset card**; it explains what each file is, how it was built, and what consumers should expect.

> **Scope:** These files summarize **note timing, density, co-note patterns, and style priors** derived from real ranked maps—not raw audio or full map zips. Downstream tools can treat them as a **compact statistical prior** for generation, analysis, or evaluation.

---

## Training context

| Item | Detail |
|------|--------|
| **Pipeline** | Python |
| **Domain** | Beat Saber **ranked** maps (standard difficulties) |
| **Scale** | **3,000+** distinct song maps worth of parsed chart data (see `training_report.json` for exact zip / map / note counts from the run that produced these files) |
| **Purpose** | Capture how ranked mappers space notes, stack simultaneous hits, and chain local patterns so generators or evaluators can stay “on distribution.” |

---

## File reference

### `ranked_spacing_profile.json`

**What it is:** A **per-difficulty spacing and density profile** built from many ranked maps.

**Contents (high level):** For difficulties such as `normal`, `expert`, and `expertplus`, you get:

- **`maps`** — how many charts contributed to that bucket  
- **`gap_p25` … `gap_p90`** — quantiles of **time gaps** between notes (beat-spacing style stats)  
- **`nps_p25` … `nps_p75`** — quantiles of **notes per second**–style density  
- **`simul_distribution`** — how often **1, 2, … simultaneous** notes appear (left/right stacks)  
- **`createdFromMaps`** — total map count feeding the profile  

**Use case:** Conditioning or validation (“does this map’s spacing look like ranked Normal / Expert+?”).

---

### `ranked_pattern_model.json`

**What it is:** The main **learned pattern / n-gram style model** (large JSON). It encodes **conditional structure** of note tokens and transitions observed across the corpus.

**Contents (high level):** Includes metadata such as:

- **`version`**, **`createdAt`**, **`source`** (input archive path used for that run)  
- **`stats`** — aggregate counts (`zips_seen`, `standard_maps`, `notes`, etc.)  
- **`global`** — **starters** and continuation statistics (tokens like `line:row:…` with `count` and probability `p`)  
- Additional sections (not fully listed here) drive **pattern continuation** from context; the file can be **very large** because it stores many n-grams / transitions.  

**Use case:** Sampling or scoring local note sequences to match ranked-map style.

---

### `brain/dataset_brain.json`

**What it is:** A **higher-level “brain”** bundle that combines **spacing priors** and the **trained pattern model** into something easier to ship to a generator or dataset consumer.

**Contents (high level):**

- **`version`** (e.g. `dataset-brain-v1`), **`createdAt`**, **`source`** description  
- **`styles`** — named **procedural style priors** (e.g. “ranked tech”, “flowy dance”, “speed map”) each with a **`vector`** of knobs (density, streams, `maxSimultaneous`, dots, walls, flow/tech bias) and a short **`description`**  
- **`retrievalIndex`** — index entries keyed by difficulty / regime (e.g. `expert`) for **retrieval-style** use alongside the vectors  

**Use case:** One file to load for “style + ranked stats + retrieval hints” without wiring every low-level JSON by hand.

---

### `training_report.json`

**What it is:** A **small JSON summary** of the training / ingestion run that produced `ranked_pattern_model.json` (and related outputs).

**Typical fields:**

- **`zips_seen`** — map archives processed  
- **`zips_without_standard_maps`** — archives skipped or without standard diffs  
- **`standard_maps`** — individual **Standard** difficulty charts parsed  
- **`notes`** — total **block / note** events counted  
- **`elapsed_sec`** — wall time for the run  
- **`model`** — path or name of the written pattern model  

**Use case:** Reproducibility, Hugging Face dataset **README stats**, or sanity checks after retraining.

---

### `training.log`

**What it is:** **Plain-text log** from the Python training run (progress + final summary).

**Contents:** Lines such as `KCODE_PROGRESS {…}` with incremental **`zips`**, **`parsed_maps`**, **`notes`**, **`elapsed_sec`**, plus a trailing **`DONE`** and optional JSON echo of final counts.

**Use case:** Debugging failed runs, comparing two trainings, or attaching evidence to a dataset card without opening huge JSON.

---

### `training.pid`

**What it is:** A single **process ID** (text file, one number) for the training job that wrote these artifacts.

**Use case:** Operational only—e.g. stopping or monitoring the process on the machine that produced the dataset. **Not** required for Hugging Face upload unless you document your local workflow.

---

## Hugging Face upload notes

1. **`ranked_pattern_model.json`** may exceed normal Git limits—use **Git LFS** or split delivery via the `datasets` library if needed.  
2. State clearly: **Python-trained** on **3,000+** Beat Saber **song maps**; artifacts are **derived statistics**, not the original maps.

---

## Directory layout

```
models/
├── README.md                   ← Hub dataset card (YAML + body)
├── ranked_spacing_profile.json
├── ranked_pattern_model.json
├── training_report.json
├── training.log
├── training.pid
└── brain/
    └── dataset_brain.json
```