网有NETHAVE
AGENT SKILL

multi-teacher-distill-lite

LinkedIn/InfoQ multi-teacher offline distillation with per-shard cache, stability switch, prune+compress (docs/MULTI_TEACHER_DISTILL.md).

License 许可
MIT
最近更新
2026-11-09
Stars
★ 1
Forks
0

01 前言 Front Matter

name
multi-teacher-distill-lite
description
LinkedIn/InfoQ multi-teacher offline distillation with per-shard cache, stability switch, prune+compress (docs/MULTI_TEACHER_DISTILL.md).

02 资源文件 1

03 SKILL.md 正文

Skill: multi-teacher-distill-lite

Follow docs/MULTI_TEACHER_DISTILL.md.

python3 scripts/multi_teacher_distill_gate.py --json

Rules:

  1. Offline first — cache teacher soft labels/embeddings; iterate the student without re-running teachers.
  2. Per-shard invalidation — never all-or-nothing teacher cache clears.
  3. Pluggable teachers — relevance | engagement | embedding | policy | click | critic.
  4. Compact student for serving — distill into skills/gates/fixtures; do not serve teacher fleets live.
  5. Stream, don’t full-stage the corpus each run.
  6. Online while exploring teachers; offline when stable (InfoQ).
  7. Compound gains — stack moderate local wins; do not default FP8 on small models.
  8. Prune + compress student serving; demand NDCG (or peer) before quality claims.
  9. No GPU multi-node training SaaS under the hard monthly cap.

04 同仓库更多 Skills

IgorGanapolsky
IgorGanapolsky/Random-Timer

writing-plans

Use when you have a spec or requirements for a multi-step task, before touching code

内容创作者写作与文案
★ 1
IgorGanapolsky
IgorGanapolsky/Random-Timer

workflow-economics-lite

Escape AI pilot purgatory: require completion-loop workflows with one owner, baseline, economic metric, system-of-record integration, and audit trail. Block orphan chatbots and interest-as-ROI claims.

Skill与智能体开发
★ 1
IgorGanapolsky
IgorGanapolsky/Random-Timer

verification-before-completion

Use when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always

★ 1
IgorGanapolsky
IgorGanapolsky/Random-Timer

value-center-lite

Apply InfoQ Rohrer value-center / VSM five questions (agency + coherence). Use when planning work, shipping releases, or auditing agent autonomy claims. Fail closed on product-only slogans without WQTU/paywall value.

★ 1
IgorGanapolsky
IgorGanapolsky/Random-Timer

using-superpowers

Use when starting any conversation - establishes how to find and use skills, requiring skill invocation before ANY response including clarifying questions

★ 1
IgorGanapolsky
IgorGanapolsky/Random-Timer

using-git-worktrees

Use when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallback

★ 1