Add orcarouter-adaptive router predictions

[ZhenghuaBao]

Method

orcarouter-adaptive is a LinUCB contextual bandit with embedding-augmented features.

• Algorithm: per-arm LinUCB ridge regression with UCB exploration. Each /decide call scores all candidate arms and picks the argmax.
• Feature vector: lexical features concatenated with a sentence-transformer embedding (all-MiniLM-L6-v2, L2-normalized, CPU inference for determinism across environments).
• Reward shape: per-query Arena Score — aligns the bandit's training objective with the leaderboard evaluation metric.
• Pre-training: per-arm parameters are initialized via offline observations before evaluation. /decide calls during evaluation are read-only on bandit state.

Pool

10 models across 5 providers:

Provider	Models
Anthropic	claude-haiku-4-5-20251001, claude-sonnet-4
Google	gemini-2.5-flash, gemini-2.5-flash-lite
OpenAI	gpt-4o-mini, gpt-5-mini
DeepSeek	deepseek-chat, deepseek-reasoner
Alibaba	qwen3-235b-a22b-instruct-2507, qwen3-30b-a3b-instruct-2507

Files

• router_inference/config/orcarouter-adaptive.json — router config
• router_inference/predictions/orcarouter-adaptive.json — full-split predictions with populated generated_result
• router_inference/predictions/orcarouter-adaptive-robustness.json — robustness predictions (routing-only)
• universal_model_names.py — provider-prefix → bare-name mappings for the pool
• model_cost/model_cost.json — token cost entries for pool models not previously listed

Local validation via router_inference/check_config_prediction_files.py passes for both full and robustness splits.

[ZhenghuaBao]

/evaluate

[github-actions]

Router Evaluation Results

Router: orcarouter-adaptive
Dataset Split: full

RouterArena Metrics

Metric	Value
RouterArena Score	0.7208
Accuracy	75.54%
Total Cost	$8.380033
Avg Cost per Query	$0.000998
Avg Cost per 1K Queries	$0.9976
Number of Queries	8400
Robustness Score	0.2262

---

Evaluation completed by RouterArena automated workflow

[ZhenghuaBao]

Hi @yl231,

All pipelines pass and evaluation results are included (full 8400 + 420 robustness splits). We'd like to merge OrcaRouter Adaptive as an online learning entry — a different regime from the offline/pre-trained routers currently on the leaderboard, intended as a reference point for contextual bandits in routing.

Thanks in advance for the review!