Data Sources

BlawkOps is built on open data. Six independent BJJ graph datasets, the Gracie Barra GB1 curriculum, and 120,279 labelled video frames from the University of Ljubljana. This page documents every source, what we extracted, and the people whose work made it possible.

811
positions (6 sources)
5,438
edges / techniques
96
GB1 techniques
120,279
labelled frames

Graph Sources

We parsed six open-source BJJ position/transition datasets into a common schema: positions (nodes), edges (transitions, submissions, sweeps, escapes, passes), and metadata. Cross-source matching uses a 31-entry alias table to normalize position names.

GrappleMap

Largest
by Eelis van der Weegen
3D position database with joint-frame representations for 596 named positions. Each transition has full skeletal animation data. The most comprehensive single-author effort to map BJJ positions spatially. We extracted 1,490 named transitions.
596 positions
1,490 edges
Public Domain

BJJ Graph

Most edges
by Diogo Seca
Structured JSON knowledge graph with 85 positions and per-perspective variants (top/bottom). Includes probability weights, decision trees, and outcome branches (success/failure/counter), producing 3,528 directional edges. Also contains the Roger Gracie Fundamental System as a tagged subsystem.
85 positions
3,528 edges
PolyForm Noncommercial 1.0

Flow-State

Belt-tagged
by iphoenix227
TypeScript-based BJJ attack decision map with gi/no-gi variants, skill level annotations (White through Black belt), priority rankings, chain families, and concept tags. Clean, opinionated, and pedagogically structured.
16 positions
175 edges
Unspecified

BJJ Graph (Clojure)

Gracie-based
by Dave Yarwood
Clojure graph built from the Gracie University Combatives and Blue Belt curriculum. 90 technique nodes with named transitions. Valuable as an independent Gracie-lineage reference that confirms our GB1 mapping from a different angle.
90 positions
206 edges
Unspecified
by Felipe Cavani
RDF/N-Quad semantic graph of BJJ positions. Bilingual (Portuguese and English). Small but structurally unique as the only Linked Data approach to BJJ positions.
12 positions
14 edges
CC BY-NC 4.0
by Ian Wessen
Minimal finite state machine representation of BJJ positions. Compact and clear, using arrow notation for one-way and bidirectional transitions. A good baseline for validating our fundamental position set.
12 positions
25 edges
MIT

Edge type distribution (5,438 total)

Transition
4,416
Sweep
418
Submission
397
Escape
171
Pass
36

Gracie Barra GB1 Curriculum

96 techniques across 8 fundamental positions, mapped to BLISP notation with phase classification (define, control, attack/defend, transition) and perspective tagging (Me/Op). 30 cross-position transitions explicitly link positions into a directed graph.

Code Position Techniques Teaching Level
STND Standing 23 Level 4 (takedowns, self-defense)
CGRD Closed Guard 22 Level 1 (first position taught)
OGRD Open Guard 14 Level 3 (sweeps, spider, guard pulling)
HGRD Half Guard 3 Level 2 (underhook sweeps)
SCTR Side Control 12 Level 0 (escapes, submissions, advancing)
MNT Mount 11 Level 0 (chokes, armbars, escapes)
TRTL Turtle 6 Level 0 (guard recovery, back take defense)
BCTR Back Control 5 Level 0 (chokes, escapes)
Total 96 + 30 cross-position transitions

Curriculum structure reconstructed from the Gracie Barra Fundamentals Program (GB1). Each technique is tagged with perspective (Me = playing the position, Op = opposing) and phase (define, control, attack/defend, transition). Teaching order follows the GB1 progression: Closed Guard first, Standing last.

ViCoS Video Dataset

120,279 labelled video frames with COCO 17-keypoint pose annotations from the University of Ljubljana's Visual Cognitive Systems Lab.

120,279
labelled frames
34
keypoints / frame
18
position classes

Hudovernik & Skocaj (2022). Video-Based Detection of Combat Positions and Automatic Scoring in Jiu-jitsu. MMSports'22, Lisbon.

CC BY-NC-SA 4.0 — University of Ljubljana, Faculty of Computer and Information Science.

Acknowledgements

BlawkOps would not exist without the open-source BJJ community. Every dataset listed here represents someone's decision to share their work freely.

Eelis van der Weegen built GrappleMap over years of solo work — 596 positions with full 3D skeletal data, released to the public domain. It remains the largest open BJJ position database.

Diogo Seca created BJJ Graph, the most edge-rich dataset we use. Its per-perspective modeling (top vs. bottom) and outcome branching (success/failure/counter) directly informed our duality and failure topology design.

Dave Yarwood mapped the Gracie University curriculum into a Clojure graph, giving us an independent Gracie-lineage reference to cross-validate against GB1.

iphoenix227 built Flow-State with belt-level skill tagging, gi/no-gi markers, and chain families — the only dataset that explicitly models pedagogical progression.

Felipe Cavani and Ian Wessen built minimal but structurally clean graphs (RDF and FSM respectively) that helped validate our fundamental position set.

Hudovernik & Skocaj at the University of Ljubljana created the ViCoS BJJ dataset — 120K labelled frames that made empirical validation of our algebraic position recognition possible.

Gracie Barra — the GB1 Fundamentals curriculum provides the pedagogical backbone for our teaching order and the 96-technique reference set.