BJJ-Bench-500: 500-Question Stratified Biomechanical, Diagnostic, Ruleset & Multi-Excerpt RAFT Evaluation Suite

BJJ-Bench-500 is a 500-question extended evaluation benchmark (100% unique prompts, 5 categories × 100 questions) for testing Brazilian Jiu-Jitsu (BJJ) and submission grappling language models and Retrieval-Augmented Generation (RAG) systems. It strictly preserves all 250 questions of maxdemarzi/bjj-bench-250 (v2) as a subset while doubling coverage across all technical domains and adding a dedicated 100-question Multi-Excerpt RAFT Synthesis category.

1. Benchmark Categories (5 x 100 = 500 Unique Questions)

Category	Count	Focus
1. Nomenclature & Lineage (nom_*)	100	Cross-lineage synonym equivalence (Inside Sankaku ↔ The Saddle / 4/11 / Honey Hole, Single-Leg X ↔ Ashi Garami, K-Guard ↔ Matrix, Cow Catcher, Whizzer, Ouchi Gari ↔ Osoto Gari, Von Flue, Paper Cutter) and canonical taxonomy classification across 163 canonical tags (457 synonyms, 45 Markov position states).
2. Diagnostic Go/No-Go Gates (diag_*)	100	10 flagship binary physical checks (Can you point your toes?, Can your heel touch your backside?, Has the elbow crossed the centerline?) + 90 stratified multi-word mechanical error callouts covering all 79 technical guides.
3. Kinematic Troubleshooting (kin_*)	100	10 flagship directional vectors & levers (down-and-back knee cut vector, perpendicular triangle angle, saddle heel exposure) + 90 highest-lever technical H2 sections covering all 79 technical guides.
4. Ruleset & Tactical Strategy (rule_*)	100	Official IBJJF / ADCC / EBI ruleset edge cases, 25 empirical Markov Position Transition / Expected-Value chaining questions (position_values), competitive counter-chain scenarios, and adversarial false-premise & safety traps.
5. Multi-Excerpt RAFT Synthesis (raft_*)	100	100 held-out 6-excerpt RAG prompts ([1]–[6], 4 oracle chunks + 2 hard-negative distractors with [Tags: ...] headers) testing 3-to-4 paragraph multi-instructor synthesis, oracle citation recall, and 0% distractor citation hallucination.

2. Schema (bjj_bench_500.jsonl)

Each JSONL line contains:

• id (string): Unique identifier (nom_*, diag_*, kin_*, rule_*, raft_*).
• category (string): One of the 5 benchmark categories above.
• question (string): The user prompt presented to the model.
• reference (string): Ground-truth reference explanation.
• rubric (object):
  • concept_groups (list[list[string]]): 3 synonym groups representing the required mechanical concepts, ruleset facts, false-premise refutations, or oracle citations (1.0 to 5.0 Kinematic Precision scale).
  • must_mention (list[string]): Primary canonical terms from each concept group.
  • penalize (list[string]): Disqualifying mechanical inversions or ruleset hallucinations.
  • forbid_regex (list[string]): Regular expressions catching sycophantic trap agreement, role inversions, or distractor citations (\[d\]).
  • oracle_indices (list[int]): Oracle excerpt numbers (1–6) for Multi-Excerpt RAFT Synthesis items ([] for non-RAFT items).
  • distractor_indices (list[int]): Hard-negative distractor excerpt numbers (1–6) for Multi-Excerpt RAFT Synthesis items ([] for non-RAFT items).

3. Live Scorecard (maxdemarzi/bjj-coach-7b-v3 on BJJ-Bench-500)

BJJ-Bench-500 Category	Questions	Kinematic Precision	Actionability	Hallucination / Anti-Distractor Pass Rate	Mean Latency
Nomenclature & Lineage	100	5.00 / 5.0	5.00 / 5.0	100.0%	4.06s
Diagnostic Go/No-Go Gates	100	5.00 / 5.0	4.99 / 5.0	100.0%	4.82s
Kinematic Troubleshooting	100	5.00 / 5.0	4.98 / 5.0	100.0%	6.18s
Ruleset & Tactical Strategy	100	5.00 / 5.0	4.78 / 5.0	100.0%	4.15s
Multi-Excerpt RAFT Synthesis	100	5.00 / 5.0	4.97 / 5.0	100.0% (0.0% distractor citations)	8.81s
Overall (BJJ-Bench-500 Full Suite)	500	5.00 / 5.0	4.94 / 5.0	100.0%	4.66s (p50)