Skip to content

Meta-eval: (7/7) Estimate meta-eval tokens and OpenRouter reference cost - #107

Open
ErlisLushtaku wants to merge 5 commits into
meta-eval/06-elo-gapfrom
meta-eval/07-cost-telemetry
Open

Meta-eval: (7/7) Estimate meta-eval tokens and OpenRouter reference cost#107
ErlisLushtaku wants to merge 5 commits into
meta-eval/06-elo-gapfrom
meta-eval/07-cost-telemetry

Conversation

@ErlisLushtaku

@ErlisLushtaku ErlisLushtaku commented Aug 15, 2026

Copy link
Copy Markdown
Collaborator

Description

Note

Packaged meta-eval split from closed PR #76. Meta-eval does not use the unified do_inference cache, and the cache stack (#94-#100) does not wire it up either: caching will be added to meta-eval once those PRs are merged.

Each judge pass records chars/4 token counts and, when a local openrouter_pricing.json is present, an estimated USD cost. Nothing is fetched at runtime. Missing pricing leaves cost null. swap_mode=both bills both passes because each orientation is its own annotation row.

This is stacked on #106.

Each judge pass records chars/4 token counts and, when a local
openrouter_pricing.json is present, an estimated USD cost. Nothing is
fetched at runtime. swap_mode=both bills both passes because each
orientation is its own annotation row.
The report dump drops None values, so a run without local OpenRouter pricing
silently lost both cost keys from results.json instead of reporting them as
null. They now carry NaN, which the artifact writer maps to JSON null the way
it already does for a degenerate run's kappa.
# Conflicts:
#	README.md
#	judgearena/benchmarks/meta_eval/annotate.py
@ErlisLushtaku ErlisLushtaku changed the title WIP: (7/7) Estimate meta-eval tokens and OpenRouter reference cost Meta-eval: (7/7) Estimate meta-eval tokens and OpenRouter reference cost Aug 17, 2026
@ErlisLushtaku
ErlisLushtaku marked this pull request as ready for review August 17, 2026 09:21
@kargibora

Copy link
Copy Markdown
Member

Each judge pass records chars/4 token counts and, when a local openrouter_pricing.json is present, an estimated USD cost. Nothing is fetched at runtime. Missing pricing leaves cost null. swap_mode=both bills both passes because each orientation is its own annotation row.

OpenRouter already provides token/reasoning counts, as well as the cost https://openrouter.ai/docs/cookbook/administration/usage-accounting

I think we should implement a class for tracking this so it gets updated for each of the pipeline and not just meta-evaluation: because we would like to also know the cost for other pairwise/elo/mt-bench pipelines.

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