AGENT SKILL

evaluation

Use rLLM evaluation APIs and CLI for AgentFlow/evaluator authoring, benchmark execution, harness selection, sandboxed eval, saved episodes, and result interpretation.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
evaluation
license
Apache 2.0
summary
Run rLLM evaluations and author AgentFlow/Evaluator protocols.
metadata
{ "disco-role": "operating" }
description
Use rLLM evaluation APIs and CLI for AgentFlow/evaluator authoring, benchmark execution, harness selection, sandboxed eval, saved episodes, and result interpretation.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

rLLM Evaluation

Use this sub-skill when the task involves rllm eval, @rllm.rollout, @rllm.evaluator, AgentFlow, Evaluator, built-in/custom harnesses, benchmark scoring, pass@k, saved episodes, or debugging an agent's interaction traces.

Start Here

  1. Read references/agentflow-and-evaluator-api.md before writing custom flows or evaluators. It captures current import locations and return-type contracts.
  2. Read references/evaluation-workflows.md for CLI recipes, provider/model selection, pass@k, sandboxed tasks, saved episode files, and programmatic run_dataset usage.
  3. Read references/troubleshooting.md for common protocol, verifier, sandbox, provider, and result-schema failures.
  4. For data layout or dataset registry work, switch to ../datasets/SKILL.md. For provider/model setup, switch to ../cli-ops/SKILL.md. For training rollouts, switch to ../training/SKILL.md.

Core Decisions

  • Is the agent a simple Python function? Use @rllm.rollout and return an Episode, Trajectory, or None. A None return is valid when gateway traces will fill steps later.
  • Does the evaluator score host-side or in a sandbox? Host-side evaluators can be fixed with --evaluator; per-task/verifier metadata is used when task config declares sandbox or script/module verifiers.
  • Does the task need a sandbox? rLLM joins the flow's needs_env, the evaluator/verifier kind, and task metadata/environment directories. Do not provision sandboxes for tasks with no consumer.
  • Is this eval-only or training data generation? Eval-only paths can save episodes/results; training paths must preserve token IDs/logprobs and route through the training sub-skill.

Safe Checks

Run the bundled protocol smoke test to verify decorators and evaluator coercion without a model call:

python scripts/agentflow_eval_smoke.py

It constructs a tiny Task, AgentConfig, Trajectory, and evaluator result; it does not call a provider or modify user state.

Avoid

  • Do not copy stale imports from old docs if they reference missing rllm.sdk modules.
  • Do not hard-code provider calls in evaluator tests; keep provider selection at the CLI/config boundary.
  • Do not use a host-side evaluator for tasks whose dataset.toml/task.toml expects sandbox-side tests unless you intentionally override that behavior.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-triton-opt

Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-solve

END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-skill

PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-pattern-fusion

Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-orchestrate

Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-feedback

Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).

★ 351