L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Développement · SKILL.md

gedi_multiclass_control

gedi_multiclass_control

Build GeDi true/false control-code inputs for multi-class and zero-shot topic control with exactly two class-conditional passes.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: gedi_multiclass_control description: Build GeDi true/false control-code inputs for multi-class and zero-shot topic control with exactly two class-conditional passes.

GeDi Multi-class Control

When to use

Use this skill when adapting GeDi to many labels or topic control without one forward pass per class. It is appropriate for AG-News-like labels and zero-shot topic words.

Inputs

  • A label or topic string.
  • Text or prompt.
  • Positive control token, default true.
  • Negative control token, default false.

Outputs

  • Positive/control sequence.
  • Negative/anti-control sequence.
  • Metadata including label, token count hint, and control_passes=2.

Workflow

  1. Normalize the label without changing its semantic content.
  2. Build true <label> <text> and false <label> <text>.
  3. During training, use true pairs for matching class/text and false pairs for nonmatching class/text.
  4. During generation, use true as desired control and false as anti-control.
  5. Record whether the label is seen or zero-shot when that information is supplied.

Validation

Run:

python tests/test_multiclass.py

Limitations

This skill formats the control-code pairs; it does not tokenize with a model or evaluate topic relevance.

Skills enregistrées