Battle Arena

Two anonymous models answer your prompt — vote for the better response. Rankings update live via Elo. Runs 100% in your browser.

vs

Try:

🎭 Demo mode: 16 simulated model personas — no server, no API keys needed. Everything (Elo, votes, favorites) is stored in your browser.

A

Model A

anonymous

VS

B

Model B

anonymous

Leaderboard

Models ranked by Elo rating from battle votes. Bars show the 95% confidence interval.

#	Model	Arena Score	95% CI	Votes	Organization	License

Elo: K=32 · ties and "both bad" count as draws · ratings seeded with priors, adjusted by real votes · CI via Wilson approximation on win-rate. Methodology inspired by lmarena/arena-rank.

Models

The catalog available for battle. Click Battle to challenge a model against a random opponent.

Battle History

Your past battles and votes (stored in your browser).

💎 Distillation

Your best-of collection — responses you saved during battles. Inspired by arena-ai's distillation idea: collect the best evaluated responses as training data for your own model.

Settings

Safety & evaluation features extracted from the Sarus Arena framework, plus data management.

🕵️ Compliance & safety

PII redaction Mask emails, phone numbers & IP addresses in model responses (like arena-ai's PII removal). Profanity guardrail Detect & censor inappropriate words in responses (like arena-ai's guardrailing). Auto-judge analysis Show formula-based evaluation + LLM-as-judge scores under each response.

📊 Evaluation methodology

AB-testing battle — side-by-side anonymous comparison (arena-ai / lmarena core)
User feedback evaluation — votes feed an Elo rating system
Formula-based evaluation — measurable response metrics: length, structure, code blocks, readability
LLM as a judge — auto-judge scores helpfulness, clarity, accuracy & formatting (heuristic simulation)
PII removal — email / phone / IP redaction
Guardrailing — profanity detection & censoring
Evaluation-based routing — "Route to winner" sends follow-ups to the top-rated model
Distillation — save the best evaluated responses as a dataset

💾 Data

Everything lives in your browser's localStorage — no server, no account, no tracking.

ℹ️ About

An educational re-implementation inspired by lmarena.ai (LMSYS Chatbot Arena), github.com/lmarena and the Sarus Arena framework (arena-ai/arena). Not affiliated with either project. All model personas are simulated.