Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add pnakhat/qa-ai-repo --skill llm-evalgit clone --depth 1 https://github.com/pnakhat/qa-ai-repoWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/pnakhat/qa-ai-repo/llm-eval)<a href="https://agentmods.dev/skills/pnakhat/qa-ai-repo/llm-eval"><img src="https://agentmods.dev/badge/skills/pnakhat/qa-ai-repo/llm-eval/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/pnakhat/qa-ai-repo/llm-eval"><img src="https://agentmods.dev/badge/skills/pnakhat/qa-ai-repo/llm-eval.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00194 | $0.03606 |
| Opus 5 | $0.00097 | $0.01803 |
| Sonnet 5 | $0.00039 | $0.00721 |
| Haiku 4.5 | $0.00019 | $0.00361 |
Grade A, and why
llm-eval scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 10d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 214 lines — stays where its author put it; the contents beside it link to each section on GitHub.
LLM Evaluation with DeepEval
Evaluating an LLM feature answers one question with numbers: "is the output good enough to ship?" You cannot answer it by reading a few outputs and nodding. The work is always in this order: build a golden dataset, pick the metric that matches the failure mode you're afraid of, set a threshold, run it as a gate, then read the score distribution. An eval with no threshold is a demo — it decorates a notebook and blocks no regression.
Two properties make LLM eval different from ordinary testing, and every guardrail here follows from them:
- The output is non-deterministic and semantic. The same input yields different valid wordings. So you assert on meaning via metrics-with-thresholds, never on exact strings — except where the output is structured (JSON, a tool call), where you go back to deterministic checks.
- The grader is often another LLM. Most quality metrics are LLM-as-judge. The judge is itself non-deterministic and can be wrong, so it must be pinned, constrained, and spot-checked against humans — a judge you never validate is a ruler you never calibrated.
The metrics — baked in
Pick by the failure mode you're guarding against, not by what's easy to compute.
Every DeepEval metric takes an LLMTestCase; the Inputs column is which fields
that metric actually reads. Direction is the trap: most metrics are maximize
(pass when score >= threshold), but hallucination, bias, and toxicity are
minimize (pass when score <= threshold).
| Metric | What it scores | Inputs (LLMTestCase fields) | Score means | Pass when |
|---|---|---|---|---|
| AnswerRelevancy | Does the output actually address the input? (generator) | input, actual_output |
relevant statements ÷ total statements in output | >= threshold |
| Faithfulness | Does the output stay true to what was retrieved? (generator) | input, actual_output, retrieval_context |
truthful claims ÷ total claims vs retrieved docs | >= threshold |
| ContextualPrecision | Are the relevant retrieved chunks ranked above noise? (retriever) | input, actual_output, expected_output, retrieval_context |
ranking-weighted relevance of retrieved nodes | >= threshold |
| ContextualRecall | Did retrieval fetch everything the answer needs? (retriever) | input, expected_output, retrieval_context |
claims in expected_output attributable to retrieval ÷ total | >= threshold |
| ContextualRelevancy | How much of what was retrieved is on-topic? (retriever noise) | input, actual_output, retrieval_context |
relevant statements in retrieval ÷ total retrieved | >= threshold |
| Hallucination | Does the output contradict known ground truth? | input, actual_output, context |
contradicted contexts ÷ total contexts | <= threshold |
| ToolCorrectness | Did the agent call the right tools? (deterministic, no judge) | input, actual_output, tools_called, expected_tools |
correctly-called tools ÷ expected (name ± args/output/order) | >= threshold |
| TaskCompletion | Did the agent accomplish the user's goal? | input, actual_output, tools_called |
judge's assessment the task's outcome was achieved | >= threshold |
| GEval (custom) | Any criterion you write in plain English | you declare evaluation_params |
chain-of-thought judge score 0–1 on your rubric | >= threshold |
| Summarization | Is the summary both accurate and complete? | input (source), actual_output |
min(alignment, coverage) | >= threshold |
| Bias | Gender/race/political/etc. bias in the output | actual_output |
share of biased opinions | <= threshold |
| Toxicity | Toxic / harmful language in the output | actual_output |
share of toxic statements | <= threshold |
What ships with it
2 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 10d ago First seen · 214 lines · 194 tokens per session scan A 92df0365aa92
llm-eval is a skill published in the GitHub repository pnakhat/qa-ai-repo (2 stars, last pushed 2mo ago), licensed MIT. It adds 194 tokens to every session and 3,606 once invoked, about $0.0010 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
cli-eval
Create and run evaluation suites, watch live benchmark progress, view scorecards, compare model performance, and integrate eval runs with CI workflows from the CLI.
model-merging
Merge multiple fine-tuned models using mergekit to combine capabilities without retraining. Use when creating specialized models by blending domain-specific expertise (math + coding + chat), improving performance beyond single models, or experimenting rapidly with model variants. Covers SLERP, TIES-Merging, DARE, Task…
darwinian-evolver
Evolve prompts/regex/SQL/code with Imbue's evolution loop.
validate
Validate Semantica pipelines, extraction quality, graph schemas, and ontology consistency. Returns structured error/warning checklists. Uses PipelineValidator, PipelineBuilder.validatepipeline(), GraphValidator, and OntologyValidator. Sub-commands: pipeline, step, dependencies, extraction, graph, ontology, performance.
launching-evals
Run, monitor, analyze, and debug LLM evaluations via nemo-evaluator-launcher. Covers running evaluations, checking status and live progress, debugging failed runs, exporting artifacts and logs, and analyzing results. ALWAYS triggers on mentions of running evaluations, checking progress, debugging failed evals…
nemo-automodel-recipe-development
Create and modify NeMo AutoModel training and evaluation recipes, including YAML structure, builders, and execution flow.