toxigen_alice_decoding
Apply ToxiGen ALICE classifier-in-the-loop weighted beam scoring for adversarial or detoxifying generation experiments.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- toxigen_alice_decoding
- description
- Apply ToxiGen ALICE classifier-in-the-loop weighted beam scoring for adversarial or detoxifying generation experiments.
02 资源文件 2
- 脚本scripts/alice_decoding.py5.0 KB
- 脚本tests/test_alice_decoding.py1.5 KB
03 SKILL.md 正文
ToxiGen ALICE Decoding
Use this skill when a task needs ALICE-style candidate ranking or a small classifier-in-the-loop decoding trace. It is suitable for reduced recovery with deterministic fake language models/classifiers and for auditing whether a full generation harness uses the paper's scoring rule.
Do not use this skill to construct demonstration prompts, compute human annotation classes, or train classifiers. It expects candidate continuations and classifier scores to be supplied by the caller.
Inputs
- Prompt text and optional current prefix.
- Candidate continuations with language-model log probabilities.
- Classifier log probabilities for
toxicandbenignclasses for each candidate. prompt_label,attackmode,lambda_l,lambda_c, and beam size.- Optional prompt-copy exclusion.
Outputs
- Selected continuation and ranked beam records.
- Per-candidate trace containing LM log probability, target class, classifier log probability, combined score, and excluded status.
- Mechanism flags for recovery: weighted scoring used, classifier target used, and prompt-copy prevention used.
Workflow
- Map the attack mode and prompt label to the classifier class to maximize.
- Remove prompt-copy candidates if configured.
- Compute
lambda_l * lm_logprob + lambda_c * classifier_logprob[target_class]. - Sort candidates by combined score and keep the requested beam size.
- Emit a trace that allows analysis to verify the mechanism independently of metric success.
Validation
Run:
python scripts/alice_decoding.py --self-test
python tests/test_alice_decoding.py
The tests check target-class mapping, classifier-dominated selection, and prompt-copy exclusion.
Limitations
This skill does not call GPT-3, HateBERT, or external APIs. In full recovery, callers should pass real model scores. In soft reduced recovery, deterministic scores are acceptable only when the result is explicitly marked as a proxy.
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).