gsm8k_recovery_evaluation
Run a bounded GSM8K verifier recovery harness with source provenance, generated-skill invocation logs, and solve-rate metrics.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- gsm8k_recovery_evaluation
- description
- Run a bounded GSM8K verifier recovery harness with source provenance, generated-skill invocation logs, and solve-rate metrics.
02 资源文件 3
- 脚本scripts/evaluation.py1.4 KB
- 脚本scripts/run_recovery.py11.1 KB
- 脚本tests/test_recovery_evaluation.py642 B
03 SKILL.md 正文
GSM8K Recovery Evaluation
Use this skill to run the soft-mode recovery experiment for the GSM8K verifier paper. It orchestrates the generated answer, candidate, verifier-training, and verifier-search skills and emits validator-compatible recovery artifacts.
Do not run this skill against the original repository during recovery. Use only the runtime handoff's allowed current-attempt snapshot files.
Inputs
- Attempt directory with
module_plan.json. - Runtime handoff with allowed snapshot paths.
- Generated skills root.
Outputs
- Recovery result with numeric
solve_rate. - Generated data item and training trace.
- Command and generated-skill invocation logs.
- Source manifest that excludes the original repository.
Workflow
- Read
environment/runtime_handoff.json. - Resolve the allowed GSM8K snapshot file.
- Generate reduced candidate records.
- Train the lightweight verifier and save loss/parameter trace.
- Rank candidates and produce predictions.
- Evaluate solve rate.
- Write mechanism checks and provenance artifacts.
Validation
Run the harness and then the Distiller recovery validator:
python scripts/run_recovery.py --attempt-dir /path/to/attempt --skills-root /path/to/generated_skills
python /share/project/yuyang/workspace/Paper2Skills/Distiller/skills/recover-paper/scripts/validate_recovery_experiment.py /path/to/attempt --output /path/to/attempt/recovery/experiment_validation.json
Limitations
This harness is a reduced proxy when full model checkpoints and GPT-3-family training infrastructure are unavailable. It must keep full-model booleans false and explicitly record reduced training as the accepted soft-mode scope.
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).