CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

write-conclusion-pack

Turn a finished local model eval into a conclusion pack (conclusions, data tables, charts, run manifest, limitations, raw data) that someone else can check and reproduce.

Quality

87%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

SKILL.md
Quality
Evals
Security

Write the Conclusion Pack

Use this after run-model-eval has produced results in a run directory (outputs/model-evals/<date>-<model>/).

The pack is what reviews, posts and vendor feedback are built from, so every claim in it must trace back to a number in the same folder.

Build the data

python3 skills/run-model-eval/holaeval.py tables --dir <run_dir>
python3 skills/run-model-eval/holaeval.py chart --dir <run_dir> --kind speed --title "<Model> on <machine>"
python3 skills/run-model-eval/holaeval.py chart --dir <run_dir> --kind capability --title "<Model> spot check"

Charts come out as SVG and, where the system can convert it, PNG (posts need PNG). If only SVG is produced, say so.

Write <run_dir>/report.md

Four sections, in the user's language:

  1. Conclusions: three to five one-line judgments, each with the condition it holds under ("on a 48GB M5 Pro, Q4 decodes at 141 tok/s"). Every one cites a number from the data section. Lead with what a reader deciding whether to use this model needs to know. Include surprises found in follow-up experiments, with both conditions of the experiment.
  2. Data: paste tables.md; reference the chart files.
  3. Run manifest: from manifest.json and the summaries: machine (chip, memory, power), engine and version, model id and file digest or quantization, sampling actually sent, the server's own defaults if they differ from the recommendation, pack id and version, samples per task, date.
  4. Limitations: one machine; samples per task; the task set is a spot check and not comparable to official benchmark scores; skipped tasks and why; anything not tested (other engines, thinking off, concurrency).

If a vendor supplied part of the task pack, label those tasks as vendor-specified wherever their results appear.

Small samples are fine as long as they are stated ("2 of 3"). Do not round a 3-sample result into a general claim.

Finish

Tell the user where the pack is, which files it contains, and the two or three findings worth acting on. Keep the raw tasks-*.jsonl and speed-*.json files in the folder so the numbers can be checked later.

Then offer to draft posts with draft-review-posts.

Repository

holaboss-ai/holaOS

Path

runtime/api-server/src/embedded-capabilities/local-model-review/skills/write-conclusion-pack/SKILL.md

Last updated
First committed

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.