evaluation-reproduction
Use PixelRAG evaluation scripts and paper reproduction harnesses safely, including reader/search serve preflights and graders.
- 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
- Use PixelRAG evaluation scripts and paper reproduction harnesses safely, including reader/search serve preflights and graders.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/eval-api-reference.md2.6 KB
- 参考文档references/paper-reproduction.md2.7 KB
- 参考文档references/troubleshooting.md2.1 KB
- 脚本scripts/pixelrag_eval_preflight.py2.4 KB
03 SKILL.md 正文
PixelRAG Evaluation and Reproduction
Use this sub-skill when the task asks to reproduce PixelRAG paper results, run eval/ benchmark cells, debug reproduce.sh, configure reader/search serves, compare retrieval modes, or understand grader/output metadata.
Start Here
- Decide whether the user wants a quick pipeline smoke or a paper-matching run.
- Quick smoke can use a public search API or a tiny self-hosted index.
- Paper Table 1 reproduction requires specific search indexes, reader model, large tile/index data, and grader credentials.
- Validate services before running a benchmark. Use pixelrag_eval_preflight.py or reproduce the shell preflight manually.
- Confirm credentials:
- OpenAI key/base URL for LLM judge on NQ/NQ-Tables and some VQA grading paths.
- Optional MiniMax key for MiniMax reader experiments.
- Record the run metadata stamped by the harness: dataset/split/n, reader, retrieval URL/status, top-k, instruction, and grader.
Read or Run
- Read paper-reproduction.md for
reproduce.sh, ports, roles, expected resources, and public-API caveats. - Read eval-api-reference.md for
run_bench.py, retrieval flags, model config, output files, and graders. - Read troubleshooting.md for empty retrieval, wrong index, judge failures, query-image path issues, and slow on-demand rendering.
- Run pixelrag_eval_preflight.py before an expensive eval.
Common Routes
| Request | Action |
|---|---|
| "Reproduce Table 1 NQ base" | Check reader Qwen/Qwen3.5-4B, base pixel serve vector count, OpenAI judge, then run the nq/base cell. |
| "Try a public API smoke" | Use direct run_bench.py --local-api-url mode; warn that public base endpoint may not match normed paper index. |
| "Why is score near zero?" | Check preflight, retrieval timeout, grader key/base URL, reader model, and whether closed-book fallback occurred. |
| "Run MiniMax reader" | Use model config rules and set MiniMax endpoint/key/context length explicitly. |
| "Need full training checkpoint eval" | Route to ../training-and-data/SKILL.md for training outputs, then return here for eval harness. |
Boundaries
- Do not download hundreds of GB of indexes or tiles unless the user explicitly approves.
- Do not launch vLLM, search serves, or long benchmark jobs silently.
- Do not treat strict exact-match NQ/NQ-Tables scores as paper numbers; the paper used an LLM judge for those cells.
- Keep benchmark outputs and reports outside the runtime skill tree.
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).