recovery_evaluation
Evaluate Tent recovery artifacts for numeric metrics, source-boundary compliance, and mechanism-faithful proxy evidence.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- recovery_evaluation
- description
- Evaluate Tent recovery artifacts for numeric metrics, source-boundary compliance, and mechanism-faithful proxy evidence.
02 资源文件 2
- 脚本scripts/evaluate_recovery.py3.1 KB
- 脚本tests/test_evaluate_recovery.py902 B
03 SKILL.md 正文
Recovery Evaluation and Mechanism Checks
Use this skill after running a Tent recovery experiment. It is appropriate for full CIFAR/ImageNet recovery and for soft-mode reduced recovery. Do not accept a proxy solely because an accuracy value is high; the trace must show target-only entropy adaptation, changed modulation parameters, and executable evidence.
Inputs
recovery_result.jsonwith paper target, metrics, commands, and mechanism checks.training_trace.jsonwith losses and parameter values before and after adaptation.source_manifest.jsonand generated skill invocation logs.module_plan.jsonfast recovery target.
Outputs
- A validation report with
ok, metric gaps, mechanism failures, and recommended decision. - Human-readable feedback for analysis artifacts.
Workflow
- Verify source manifests do not include original repository paths.
- Confirm recovery target metadata matches the module plan target.
- Check that at least one numeric metric is present.
- For proxy runs, require mechanism checks for target-only adaptation, entropy loss, normalization-affine update, and optimizer execution.
- Compare
loss_beforeandloss_after, plusparams_beforeandparams_after. - Recommend
acceptonly when executable evidence and mechanism checks pass.
Validation
Run python scripts/evaluate_recovery.py --self-test. The self-test checks a valid proxy trace and an invalid unchanged-parameter trace.
Limitations
This evaluator complements, but does not replace, Distiller's recovery experiment validator. Always run both in final recovery.
Refinement Cycle 4
An unchanged-parameter ablation produced a refine decision; proxy acceptance must require parameter or optimizer-state change, not metric text alone.
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).