toxigen_classifier_recovery
Run a bounded ToxiGen-style classifier update and report validator-compatible before-after loss, AUC, and parameter evidence.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- toxigen_classifier_recovery
- description
- Run a bounded ToxiGen-style classifier update and report validator-compatible before-after loss, AUC, and parameter evidence.
02 资源文件 2
- 脚本scripts/classifier_recovery.py6.4 KB
- 脚本tests/test_classifier_recovery.py1.3 KB
03 SKILL.md 正文
ToxiGen Classifier Recovery
Use this skill when a recovery experiment needs to demonstrate the paper's classifier-improvement mechanism without full HateBERT or RoBERTa fine-tuning. It trains a tiny logistic classifier over deterministic text features and emits before/after metrics.
Do not use this skill to claim full paper Table 4 reproduction. A result from this skill is a reduced proxy unless it is replaced by a real model fine-tuning stack.
Inputs
- Generated or curated examples with
textandlabel. - Optional group and generation method metadata.
- Learning rate and update-step count.
Outputs
loss_before,loss_after,auc_before,auc_after.params_before,params_after, andoptimizer_state_changed.- Per-example probabilities before and after the update.
Workflow
- Extract deterministic features from each text: bias/toxicity cue count, benign cue count, identity cue count, and normalized length.
- Initialize logistic weights and bias to zero unless supplied.
- Compute binary cross-entropy and pairwise AUC.
- Run bounded gradient-descent updates.
- Recompute metrics and emit a trace with parameter-change evidence.
Validation
Run:
python scripts/classifier_recovery.py --self-test
python tests/test_classifier_recovery.py
The tests verify that parameters change and that loss does not increase on a simple fixture.
Limitations
This skill is intentionally small and deterministic. It is useful for mechanism-faithful soft recovery, edge-case tests, and harness validation, not for reporting deployed toxicity classifier performance.
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).