Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 8 additions & 0 deletions model_cost/model_cost.json
Original file line number Diff line number Diff line change
Expand Up @@ -338,5 +338,13 @@
"x-ai/grok-4.3": {
"input_token_price_per_million": 1.25,
"output_token_price_per_million": 2.5
},
"gpt-oss-20b": {
"input_token_price_per_million": 0.03,
"output_token_price_per_million": 0.13
},
"xiaomi/mimo-v2.5": {
"input_token_price_per_million": 0.14,
"output_token_price_per_million": 0.28
}
}
18 changes: 18 additions & 0 deletions router_inference/config/sqwish-fvcta8.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,18 @@
{
"pipeline_params": {
"router_name": "sqwish-fvcta8",
"router_cls_name": "SqwishRouter",
"models": [
"deepseek/deepseek-v4-flash",
"gemini-3.1-pro-preview",
"google/gemini-3.1-flash-lite",
"gpt-5.4-mini",
"gpt-oss-120b",
"gpt-oss-20b",
"qwen/qwen3-235b-a22b-2507",
"qwen/qwen3-next-80b-a3b-instruct",
"xiaomi/mimo-v2.5"
],
"description": "Sqwish Router jointly selects among 16 model-budget actions over nine models. A 10-member ensemble predicts per-action quality from task features and full model fingerprints, then applies a cost-aware utility rule with an uncertainty penalty. A separately trained low-success classifier, fitted only on Sqwish's offline synthetic generation corpus and no RouterArena answers, model outcomes, scores or labels, diverts predicted-unsolved tasks to the cheap backup arm gpt-oss-20b when that arm is admissible. The same frozen selector, diversion rule and admissibility mask are applied to the full and robustness splits. Budget is a first-class routing action implemented through an ordinal generation-control lever covering reasoning effort, verbosity, answer style and output cap; its full cost is captured by the reported token usage. No RouterArena answers, model outcomes, scores or labels were used to fit or tune the router."
}
}
1 change: 1 addition & 0 deletions router_inference/predictions/sqwish-fvcta8-robustness.json

Large diffs are not rendered by default.

Loading
Loading