Skip to content

Add chuzom-solo-v32: cost-aware single-model baseline (deepseek-v3.2)#161

Open
ypollak2 wants to merge 2 commits into
RouteWorks:mainfrom
ypollak2:submit/chuzom-solo-v32
Open

Add chuzom-solo-v32: cost-aware single-model baseline (deepseek-v3.2)#161
ypollak2 wants to merge 2 commits into
RouteWorks:mainfrom
ypollak2:submit/chuzom-solo-v32

Conversation

@ypollak2

@ypollak2 ypollak2 commented Jul 7, 2026

Copy link
Copy Markdown
Contributor

Summary

Router: chuzom-solo-v32 — a deliberately simple, fully-clean cost-aware baseline that routes every query to deepseek/deepseek-v3.2, selected for best accuracy-per-dollar from published benchmarks.

Compliance

  • No RouterArena-derived supervision — no per-dataset accuracy, judge scores, or oracle entries used anywhere. Model choice is justified by public benchmark reports only.
  • No benchmark/prompt-template fingerprinting — this is model-selection only.
  • Shared evaluator UNMODIFIEDllm_evaluation/metrics.py and router_inference/compare_router_accuracy.py are byte-identical to main (this PR's diff touches only the config, prediction files, and the new router class).

Files

  • router_inference/config/chuzom-solo-v32.json
  • router_inference/predictions/chuzom-solo-v32.json (full, 8400)
  • router_inference/predictions/chuzom-solo-v32-robustness.json (420)
  • router_inference/router/chuzom_solo_v32.py

Test plan

  • check_config_prediction_files.py chuzom-solo-v32 full → all checks pass
  • Trigger /evaluate

🤖 Generated with Claude Code

Routes every query to deepseek-v3.2, selected for best accuracy-per-dollar
from PUBLISHED benchmarks. A deliberately simple, fully-clean baseline:

- No RouterArena-derived supervision (no per-dataset accuracy, judge scores,
  or oracle entries used anywhere).
- No benchmark/prompt-template fingerprinting; routing is model-selection only.
- Shared evaluator UNMODIFIED (llm_evaluation/metrics.py and
  router_inference/compare_router_accuracy.py identical to main).

Files: config, full + robustness prediction files, and the router class.
@ypollak2

ypollak2 commented Jul 7, 2026

Copy link
Copy Markdown
Contributor Author

/evaluate

@ypollak2

ypollak2 commented Jul 7, 2026

Copy link
Copy Markdown
Contributor Author

/evaluate

@github-actions

github-actions Bot commented Jul 7, 2026

Copy link
Copy Markdown

Router Evaluation Results

Router: chuzom-solo-v32
Dataset Split: full

RouterArena Metrics

Metric Value
RouterArena Score 0.7061
Accuracy 70.59%
Total Cost $0.844211
Avg Cost per Query $0.000101
Avg Cost per 1K Queries $0.1005
Number of Queries 8400
Abnormal Entries 0
Robustness Score 1.0000

Evaluation completed by RouterArena automated workflow

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.

1 participant