🚀 PQN.AI – Quantum-Inspired LLM Submission for Chatbot Arena

API Access (OpenAI-compatible)

Base URL: https://api.pqn.ai/v1

Auth: Authorization: Bearer

Models: pqn-ai/pqn-ai-v1

Rate limit: 30 req/min (pilot)

Context length: 32,768 tokens (Max output: 2,048 tokens)

Throughput (A100 80G): ~110 tok/s @ fp16

Tokenizer: tiktoken (gpt-4o compatible)

Reproducible Evals

Repo: P-QUANTUM
(tag: arena-eval-2025-09-17)

Script: scripts/run_arena_evals.sh (MMLU / ARC / TruthfulQA / HellaSwag)

Seed: 42 (fully reproducible)

Logs & Results: results/arena_2025-09-17/*.jsonl + aggregate.json

Hardware: A100 80G ×1, CUDA 12.1

Aggregate Scores (pilot run):

MMLU: 82.0

ARC-Challenge: 90.0

TruthfulQA: 90.2

HellaSwag: 85.5

Overall: Competitive with GPT-4-level baselines, reproducible with seed 42.

Safety

Output moderation: OpenAI-style heuristics + blocklists (hate / sexual / violence / self-harm)

Refusal policies: Documented in safety/PQN_Safety_Policy.md

License

Code: MIT

Weights: MIT (non-viral)

Contact

Email: team@pqn.ai

Demo: HF Space – PQN.AI Demo

(maintainers: queue disabled, unlimited access for testing)

Temporary API Key

If you need a temporary API key to validate, please share a maintainer email and we’ll DM a key.

The issue provides API, evaluation, safety, and demo links, but names no FastChat file, test, or entry point. First determine whether FastChat has a documented model-submission or Chatbot Arena integration path, then identify the required project changes and validation steps. Done should mean the submission is integrated and its evaluation or serving behavior is verified.