evaluation-reproduction
Route Qwen benchmark reproduction and evaluation-script selection for C-Eval, MMLU, CMMLU, GSM8K, HumanEval, and plugin workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- evaluation-reproduction
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Route Qwen benchmark reproduction and evaluation-script selection for C-Eval, MMLU, CMMLU, GSM8K, HumanEval, and plugin workflows.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/benchmark-reproduction.md2.3 KB
- 参考文档references/evaluation-script-reference.md2.3 KB
- 参考文档references/troubleshooting.md1.2 KB
- 脚本scripts/qwen_eval_command_builder.py2.0 KB
03 SKILL.md 正文
Qwen Evaluation and Reproduction
Use this sub-skill when the user wants to reproduce Qwen benchmarks, select an evaluation script, prepare data directories, or understand the differences between base and chat evaluation paths.
Safe start
- Use
scripts/qwen_eval_command_builder.pyto print a benchmark command plan without downloading datasets or loading a model. - Treat HumanEval execution as sandbox-sensitive. Do not run untrusted generated code without explicit approval.
- Do not assume the chat evaluation scripts reproduce the same protocol as external benchmark harnesses; the repository documentation explicitly notes 0-shot chat results in some places where 5-shot results came from another system.
Routes
| User request | Read |
|---|---|
| C-Eval, MMLU, CMMLU, GSM8K, HumanEval, or plugin benchmark overview and file layout | references/benchmark-reproduction.md |
| Script flags, model/checkpoint selection, dataset arguments, and base-vs-chat command differences | references/evaluation-script-reference.md |
| Missing dataset, wrong result path, sandbox concern, or benchmark dependency issue | references/troubleshooting.md |
Boundaries
- For inference of the underlying checkpoint, use
../inference-model-loading/SKILL.md. - For prompt/tool-use behavior, use
../prompting-tool-use-tokenization/SKILL.md. - For serving or deployment, use
../serving-deployment/SKILL.md. - For finetuning or quantization, use
../finetuning-quantization/SKILL.md.
Operating rules
- Benchmark commands are only useful when the dataset layout, checkpoint family, and seed are explicit.
- Never present a score without the benchmark dataset and result file locations the repo expects.
- Keep the repository's warning about HumanEval untrusted code execution visible.
04 同仓库更多 Skills
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.
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.
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.
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.
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.
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).