Skip to content

Add OVA (ova-debate) as API-based model endpoint for Chatbot Arena#3781

Open
jacobrlutz wants to merge 1 commit intolm-sys:mainfrom
jacobrlutz:add-ova-debate-model
Open

Add OVA (ova-debate) as API-based model endpoint for Chatbot Arena#3781
jacobrlutz wants to merge 1 commit intolm-sys:mainfrom
jacobrlutz:add-ova-debate-model

Conversation

@jacobrlutz
Copy link

Add OVA (ova-debate) — API-based meta-model endpoint

What is OVA?

OVA is a confidence-adaptive multi-model debate system. Instead of calling a single LLM, it runs a structured debate across Claude, GPT, Grok, Gemini, and Perplexity, then synthesises a single final answer. The depth of reasoning scales with question difficulty:

  • 3 rounds — initial drafts → mutual critique → synthesis (standard queries)
  • 5 rounds — adds a vote round and a winner's definitive synthesis
  • 7–9 rounds — adds axiom decomposition, frame-breaking (null-hypothesis mode), and first-principles escalation for hard, ambiguous, or previously unresolved problems

OVA is publicly available as an OpenAI-compatible API.

Integration Method

Method 1 — third-party self-hosted API

The endpoint is fully OpenAI-compatible (api_type: "openai"), so no custom code is needed in api_provider.py.

Files changed

  1. api_endpoints.json — new entry for ova-debate
  2. fastchat/model/model_registry.py — new register_model_info call

Endpoint details

Field Value
api_base https://ovamind-production.up.railway.app/api/v1
model_name ova-debate
api_type openai
temperature 0.3
text-arena true
vision-arena false

The API key will be shared privately with the Arena team — please reach out and we will send it directly.

Availability commitment

The endpoint is hosted on Railway (persistent, no sleep/spin-down). OVA commits to keeping the endpoint live for a minimum of 90 days from Arena listing, with advance notice of any planned deprecation.

Latency note

OVA's 3-round fast path targets sub-30 seconds for typical queries (4096 token cap when called via the arena key). We are aware this is slower than single-model systems and believe the quality improvement from multi-agent debate justifies the latency.

Contact

OVA is a confidence-adaptive multi-model debate system that orchestrates
Claude, GPT, Grok, Gemini, and Perplexity in structured debate rounds before
synthesizing a single final answer. The endpoint is fully OpenAI-compatible.

Changes:
- api_endpoints.json: new ova-debate entry (api_type: openai)
- fastchat/model/model_registry.py: new register_model_info entry

Endpoint: https://ovamind-production.up.railway.app/api/v1
Docs: https://github.com/OVA-Mind-LLC/ovamind
Made-with: Cursor
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants