|
| 1 | +--- |
| 2 | +name: arena |
| 3 | +description: Trinity Arena 2.0 — LLM Battle Platform dashboard. Live leaderboard, battle history, scientific metrics, deploy to Railway. |
| 4 | +user_invocable: true |
| 5 | +--- |
| 6 | + |
| 7 | +# Trinity Arena 2.0 — LLM Battle Dashboard |
| 8 | + |
| 9 | +## Overview |
| 10 | + |
| 11 | +Arena is a pure-Zig LLM battle platform with ELO rankings (LMSYS-style). |
| 12 | +Binary: `zig-out/bin/arena` | Source: `src/arena/` (7 files, ~1200 LOC) |
| 13 | + |
| 14 | +## Step 1: Collect Arena State |
| 15 | + |
| 16 | +```bash |
| 17 | +# Check if arena binary exists |
| 18 | +test -f zig-out/bin/arena && echo "ARENA_BIN:OK" || echo "ARENA_BIN:MISSING" |
| 19 | + |
| 20 | +# Check leaderboard data |
| 21 | +cat data/arena/leaderboard.json 2>/dev/null || echo "LEADERBOARD:EMPTY" |
| 22 | + |
| 23 | +# Count battle results |
| 24 | +wc -l data/arena/arena_results.jsonl 2>/dev/null || echo "BATTLES:0" |
| 25 | + |
| 26 | +# Last 5 battles |
| 27 | +tail -5 data/arena/arena_results.jsonl 2>/dev/null |
| 28 | + |
| 29 | +# Check arena server running |
| 30 | +lsof -ti:8080 2>/dev/null && echo "SERVER:UP" || echo "SERVER:DOWN" |
| 31 | + |
| 32 | +# Task catalog size |
| 33 | +grep -c '.id =' src/arena/tasks.zig 2>/dev/null || echo "0" |
| 34 | + |
| 35 | +# Fighter kinds available |
| 36 | +grep -c 'FighterKind' src/arena/types.zig 2>/dev/null || echo "0" |
| 37 | + |
| 38 | +# Check Railway deployment |
| 39 | +cat .trinity/arena_deploy.json 2>/dev/null || echo "DEPLOY:NONE" |
| 40 | +``` |
| 41 | + |
| 42 | +## Step 2: Render Dashboard |
| 43 | + |
| 44 | +``` |
| 45 | +⚔ TRINITY ARENA 2.0 — LLM Battle Platform |
| 46 | +══════════════════════════════════════════════════ |
| 47 | +
|
| 48 | +📊 НАУЧНАЯ БАЗА |
| 49 | +┌─────────────────────────────────────────────────┐ |
| 50 | +│ ELO System: LMSYS Chatbot Arena compatible │ |
| 51 | +│ Formula: E = 1/(1+10^((Rb-Ra)/400)), K=32 │ |
| 52 | +│ Judge: LLM-as-judge (Anthropic/OpenAI) │ |
| 53 | +│ Debiasing: length-bias correction (WildBench) │ |
| 54 | +│ Win strength: much_better / slightly_better │ |
| 55 | +│ Reference: Zheng et al. 2023 "LMSYS Arena" │ |
| 56 | +│ Li et al. 2024 "WildBench" │ |
| 57 | +└─────────────────────────────────────────────────┘ |
| 58 | +
|
| 59 | +🏆 LEADERBOARD |
| 60 | +┌──────────────────┬──────┬─────┬─────┬─────┬───────┐ |
| 61 | +│ Fighter │ ELO │ W │ L │ T │ Total │ |
| 62 | +├──────────────────┼──────┼─────┼─────┼─────┼───────┤ |
| 63 | +│ {from leaderboard.json, sorted by ELO desc} │ |
| 64 | +└──────────────────┴──────┴─────┴─────┴─────┴───────┘ |
| 65 | +
|
| 66 | +📋 TASK CATALOG: {N} tasks |
| 67 | + math: 7 | coding: 7 | reasoning: 6 |
| 68 | + Difficulty: easy/medium/hard |
| 69 | +
|
| 70 | +🥊 RECENT BATTLES (last 5) |
| 71 | + {from arena_results.jsonl} |
| 72 | +
|
| 73 | +⚙ INFRASTRUCTURE |
| 74 | + Binary: {OK/MISSING} (zig-out/bin/arena) |
| 75 | + Server: {UP/DOWN} (:8080) |
| 76 | + Cloud: {DEPLOYED/NOT DEPLOYED} (Railway) |
| 77 | + Data: data/arena/ |
| 78 | +
|
| 79 | +🔬 НАУЧНЫЕ МЕТРИКИ |
| 80 | + Battle convergence: {total battles needed for stable ELO ≈ 30 per pair} |
| 81 | + Coverage: {pairs tested / total possible pairs} |
| 82 | + Judge agreement: {if multiple judges — inter-annotator κ} |
| 83 | + Length-bias corrections: {count from results} |
| 84 | +``` |
| 85 | + |
| 86 | +## Step 3: Quick Actions |
| 87 | + |
| 88 | +Based on state, suggest actions: |
| 89 | + |
| 90 | +| Condition | Action | |
| 91 | +|-----------|--------| |
| 92 | +| ARENA_BIN:MISSING | `zig build arena` | |
| 93 | +| SERVER:DOWN | `./zig-out/bin/arena serve &` | |
| 94 | +| LEADERBOARD:EMPTY | `./zig-out/bin/arena battle "2+2" --a echo --b echo` | |
| 95 | +| DEPLOY:NONE | Deploy to Railway (see Step 4) | |
| 96 | +| < 30 battles per pair | "Need more battles for stable ELO" | |
| 97 | + |
| 98 | +Print 2-3 concrete commands the user can run. |
| 99 | + |
| 100 | +## Step 4: Cloud Deployment |
| 101 | + |
| 102 | +When user asks to deploy Arena to Railway cloud: |
| 103 | + |
| 104 | +### Dockerfile |
| 105 | + |
| 106 | +Create `deploy/Dockerfile.arena`: |
| 107 | +```dockerfile |
| 108 | +FROM debian:bookworm-slim AS build |
| 109 | +RUN apt-get update && apt-get install -y curl xz-utils && \ |
| 110 | + curl -L https://ziglang.org/download/0.15.2/zig-linux-x86_64-0.15.2.tar.xz | tar -xJ -C /opt && \ |
| 111 | + ln -s /opt/zig-linux-x86_64-0.15.2/zig /usr/local/bin/zig |
| 112 | +WORKDIR /app |
| 113 | +COPY . . |
| 114 | +RUN zig build arena |
| 115 | + |
| 116 | +FROM debian:bookworm-slim |
| 117 | +COPY --from=build /app/zig-out/bin/arena /usr/local/bin/arena |
| 118 | +RUN mkdir -p /data/arena |
| 119 | +ENV ARENA_PORT=8080 |
| 120 | +EXPOSE 8080 |
| 121 | +ENTRYPOINT ["/usr/local/bin/arena", "serve"] |
| 122 | +``` |
| 123 | + |
| 124 | +### Railway Deploy Commands |
| 125 | + |
| 126 | +```bash |
| 127 | +source .env |
| 128 | + |
| 129 | +# Create service |
| 130 | +ARENA_SVC=$(curl -s https://backboard.railway.app/graphql/v2 \ |
| 131 | + -H "Authorization: Bearer $RAILWAY_API_TOKEN" \ |
| 132 | + -H "Content-Type: application/json" \ |
| 133 | + -d '{"query":"mutation{serviceCreate(input:{name:\"trinity-arena\",projectId:\"aa0efa7f-95e6-4466-8de6-43945a031365\"}){id}}"}' \ |
| 134 | + | python3 -c "import sys,json; print(json.load(sys.stdin)['data']['serviceCreate']['id'])") |
| 135 | + |
| 136 | +echo "Arena service: $ARENA_SVC" |
| 137 | + |
| 138 | +# Set config: builder=NIXPACKS won't work, must set dockerfilePath |
| 139 | +curl -s https://backboard.railway.app/graphql/v2 \ |
| 140 | + -H "Authorization: Bearer $RAILWAY_API_TOKEN" \ |
| 141 | + -H "Content-Type: application/json" \ |
| 142 | + -d "{\"query\":\"mutation{serviceInstanceUpdate(input:{serviceId:\\\"$ARENA_SVC\\\",environmentId:\\\"6748f1ad-9c2f-4b71-9a90-67f40ce34dc9\\\",source:{image:\\\"ghcr.io/ghashtag/trinity-arena:latest\\\"}})}\"}\"}" |
| 143 | + |
| 144 | +# Save deploy info |
| 145 | +echo "{\"service_id\":\"$ARENA_SVC\",\"deployed_at\":\"$(date -u +%Y-%m-%dT%H:%M:%SZ)\",\"url\":\"pending\"}" > .trinity/arena_deploy.json |
| 146 | +``` |
| 147 | + |
| 148 | +### Docker Build & Push |
| 149 | + |
| 150 | +```bash |
| 151 | +docker build -f deploy/Dockerfile.arena -t ghcr.io/ghashtag/trinity-arena:latest . |
| 152 | +docker push ghcr.io/ghashtag/trinity-arena:latest |
| 153 | +``` |
| 154 | + |
| 155 | +## Step 5: CLI Reference |
| 156 | + |
| 157 | +All arena commands (run via `./zig-out/bin/arena` or `tri arena`): |
| 158 | + |
| 159 | +| Command | Description | |
| 160 | +|---------|-------------| |
| 161 | +| `arena serve` | Start HTTP server on :8080 | |
| 162 | +| `arena battle <prompt>` | Run CLI battle (default: trinity-hslm vs echo) | |
| 163 | +| `arena battle "X" --a gpt-4o --b claude-sonnet --judge` | Battle with auto-judge | |
| 164 | +| `arena leaderboard` | Show ELO rankings | |
| 165 | +| `arena bench math` | Run all math tasks | |
| 166 | +| `arena bench all` | Run all categories | |
| 167 | +| `arena tasks` | List task catalog | |
| 168 | +| `arena register <name> <kind> [model]` | Register new fighter | |
| 169 | + |
| 170 | +## Step 6: HTTP API Reference |
| 171 | + |
| 172 | +| Endpoint | Method | Description | |
| 173 | +|----------|--------|-------------| |
| 174 | +| `/battle` | POST | Create battle: `{"prompt":"...","fighter_a":"...","fighter_b":"..."}` | |
| 175 | +| `/battle/:id` | GET | Get battle status | |
| 176 | +| `/leaderboard` | GET | Current ELO rankings JSON | |
| 177 | +| `/tasks` | GET | Task catalog JSON | |
| 178 | +| `/battle/:id/vote` | POST | Submit manual vote | |
| 179 | + |
| 180 | +## Scientific References |
| 181 | + |
| 182 | +- **ELO Rating**: Elo, A. (1978). "The Rating of Chessplayers, Past and Present" |
| 183 | +- **LMSYS Arena**: Zheng et al. (2023). "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena" |
| 184 | +- **WildBench**: Lin et al. (2024). "WildBench: Benchmarking Language Models with Challenging Tasks from Real Users" |
| 185 | +- **Length Bias**: Wang et al. (2024). "Large Language Models are not Fair Evaluators" — verbosity bias in LLM judges |
| 186 | +- **K-factor**: K=32 (same as LMSYS default, chess rapid); higher K = faster convergence, more volatile |
| 187 | +- **Bradley-Terry**: Arena ELO is equivalent to Bradley-Terry model coefficients when fitted via MLE |
| 188 | +- **Bootstrap CI**: For confidence intervals, resample battles 1000x and recompute ELO (not implemented yet) |
0 commit comments