Babysitter setfit-few-shot
SetFit few-shot learning for efficient intent classification with minimal data
install
source · Clone the upstream repo
git clone https://github.com/a5c-ai/babysitter
Claude Code · Install into ~/.claude/skills/
T=$(mktemp -d) && git clone --depth=1 https://github.com/a5c-ai/babysitter "$T" && mkdir -p ~/.claude/skills && cp -r "$T/library/specializations/ai-agents-conversational/skills/setfit-few-shot" ~/.claude/skills/a5c-ai-babysitter-setfit-few-shot && rm -rf "$T"
manifest:
library/specializations/ai-agents-conversational/skills/setfit-few-shot/SKILL.mdsource content
SetFit Few-Shot Skill
Capabilities
- Train SetFit models with few examples per class
- Configure contrastive learning settings
- Implement efficient classification pipelines
- Design few-shot training strategies
- Set up model evaluation
- Deploy lightweight classifiers
Target Processes
- intent-classification-system
Implementation Details
SetFit Advantages
- Few Examples: 8-16 examples per class
- No Prompts: No prompt engineering needed
- Fast Training: Minutes vs hours
- Small Models: Sentence transformer base
Training Process
- Contrastive fine-tuning of embeddings
- Classification head training
- Iterative sampling strategies
Configuration Options
- Base sentence transformer model
- Number of training examples
- Contrastive learning epochs
- Classification head architecture
- Evaluation metrics
Best Practices
- Diverse few-shot examples
- Balance class examples
- Use appropriate base model
- Validate on held-out data
Dependencies
- setfit
- sentence-transformers