arm
se_gemma_l_n100_v1 baseline
One config on summeval_100, scored one way; its repeats averaged per item.
- dataset
- summeval_100
- scored against
- gold/summeval_100 gold
- goal
- A scorer that agrees with experts on summaries (SummEval)
- runs
- 1
- headline
- tau 0.538
- last run
- 2026-10-18 09:52
01metrics
| metric | mean | items | better |
|---|---|---|---|
| tau | 0.538 | 100 | higher |
| cost_usd | $0.000278 | 100 | lower |
| latency_ms | 6670.3 | 100 | lower |
| parsed | 1.000 | 100 | neutral |
| reasoning_tokens | 0.000 | 100 | neutral |
| rho | 0.598 | 100 | higher |
| rho_coherence | 0.659 | 100 | higher |
| rho_consistency | 0.638 | 100 | higher |
| rho_fluency | 0.518 | 100 | higher |
| rho_relevance | 0.575 | 100 | higher |
| tau_coherence | 0.566 | 100 | higher |
| tau_consistency | 0.614 | 100 | higher |
| tau_fluency | 0.477 | 100 | higher |
| tau_relevance | 0.497 | 100 | higher |
| tie_rate | 0.405 | 100 | neutral |
| tokens_in | 2034.0 | 100 | lower |
| tokens_out | 217.5 | 100 | lower |
| truncated | 0.000 | 100 | neutral |
02cost and speed
- cost per item
- $0.000278 · 2,251 tok
- speed per item
- p50 6,204 · p95 10,768 ms
Per item, over its items: cost as dollars and tokens (a local model is free but still uses tokens), speed as the median and the slow tail (p95). A goal that names this arm weighs them in its decision.
03runs
| run | made | items | cost | how |
|---|---|---|---|---|
se_gemma_l_n100_v1_20261018T094137Z | 2026-10-18 09:52 | 100 | $0.027841 | measured |
04config
config se_gemma_l_n100_v1 · adapter examples/summeval-judging/adapter.py · credentials are never shown
- family
gemma- max_tokens
1500- model
eval-gemma-31b- price_tier
large- prompt_file
prompt.txt- provider
litellm- reasoning
{"enabled": false}- temperature
0.0- usd_per_mtok_in
0.1- usd_per_mtok_out
0.2
05deploy sheet
- measured by
se_gemma_l_n100_v1_20261018T094137Zon summeval_100- against
- gold/summeval_100
- model sent
eval-gemma-31b- model record
gemma4-31b- deployment
- hosted · homelab LiteLLM (the fleet's gateway) · eval-gemma-31b
- server
- not recorded
- build
- none recorded (a hosted model has no digest)
- settings sent
max_tokens: 1500model: eval-gemma-31bprompt_file: prompt.txtprovider: litellmreasoning: {"enabled": false}temperature: 0.0- not sent
- bookkeeping:
family, price_tier, usd_per_mtok_in, usd_per_mtok_out
the record setsreasoning_effort: none; this arm did not send it - declared
settings.prompt_sha256: 6c02d00929048d9c3fea511af40c3ddf90fdf1550afb7713d3f5f4d86e44f5eesettings.reasoning: {'enabled': False}settings.temperature: 0.0deployment.alias: eval-gemma-31bdeployment.endpoint: http://homelab:4001deployment.id: openrouter/google/gemma-4-31b-itdeployment.id_source: proxy-model-infodeployment.provider_routing: None- reported per item
finish_reason: stopprovider: CoreWeave, DeepInfraresponse_model: eval-gemma-31b- adapter
examples/summeval-judging/adapter.pyb757ca07e338- harness
- 90b88fd29bd7
No capture here says what production ran (the adapter's production_ran).
Command: evh deploy-sheet --config se_gemma_l_n100_v1 [--capture C]
06conditions
- data.items_sha256
b59999b281b2cbe87c57e4af34cc475c01ba1d531c7b77cb93a136665119c613- deployment.model
eval-gemma-31b- deployment.source
adapter- deployment.upstream
openrouter/google/gemma-4-31b-it- fingerprint.version
5- library.numpy
1.26.4- library.openai
3.31.0- library.rouge-score
0.1.2- library.torch
2.14.1- library.transformers
4.55.4- machine.cpu_count
32- machine.label
omarchy- machine.platform
Linux/x86_64- machine.python
3.11.16- machine.release
7.2.5-3-omarchy- reference.sha256
7b5e618e23b7f2b222665d83e99006eb0e9a51066e63e2e5ff8c51ae3c696c21- scorer.id
examples/summeval-judging/adapter.py- scorer.sha256
b757ca07e3388b96f6eeefc7033c421ca6b74ae836785f683787dd9edc870473
07rounds
- Round r1 of A scorer that agrees with experts on summaries (SummEval), closed: one of its arms.