direct_rlaif_rewarding
Compute direct-RLAIF rewards from one-to-ten score-token logits without training a separate reward model.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- direct_rlaif_rewarding
- description
- Compute direct-RLAIF rewards from one-to-ten score-token logits without training a separate reward model.
02 资源文件 2
- 脚本scripts/direct_reward.py3.6 KB
- 脚本tests/test_direct_reward.py1.8 KB
03 SKILL.md 正文
Direct RLAIF Rewarding
Use this skill when a RLAIF experiment needs the paper's direct reward path: a scorer rates one generated response with score tokens 1 through 10, the token logits become a score distribution, and the expected score is mapped to a normalized reward.
Do not use this skill for pairwise preference labeling or reward-model distillation. Direct-RLAIF is a single-response reward mechanism.
Inputs
task:summarizationorhelpful_dialogue.context: source post or dialogue history.response: generated text to score.- A scorer that returns logits for score tokens
1through10.
Outputs
score_probabilities: probabilities for scores 1 through 10.expected_score: the probability-weighted score.normalized_reward: expected score mapped from[1, 10]to[-1, 1].- Prompt metadata for auditing.
Workflow
- Build the direct scoring prompt for the task and candidate response.
- Obtain logits for score tokens
1through10. - Apply stable softmax over score-token logits.
- Compute the expected score.
- Normalize the reward with
((expected_score - 1) / 9) * 2 - 1. - Check the reward remains inside
[-1, 1].
Validation
Run:
python scripts/direct_reward.py --smoke
python tests/test_direct_reward.py
The tests verify normalization, monotonic score ordering, uniform-distribution behavior, and output bounds.
Limitations
The bundled scorer is deterministic and is meant for reduced recovery only. A full d-RLAIF run needs an actual off-the-shelf LLM with score-token likelihoods during policy training.
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).