feat(skills): Create anti-hallucination-training skill for calibration and refusal data

[akaszubski]

Summary

Create a skill documenting anti-hallucination training data generation and calibration workflows.

Context

ReAlign has AntiHallucinationGenerator and calibration training capabilities but they're not documented in skills.

Implementation Approach

Create .claude/skills/anti-hallucination-training.md documenting:

Anti-Hallucination Data Types

1. Appropriate Refusal

   • Refuse when asked about unknown facts
   • Example: "I don't have information about events after my training cutoff"

2. Hedged Uncertainty

   • Express uncertainty for partial knowledge
   • Example: "I'm not certain, but I believe..."

3. Confident Answers

   • Answer confidently for verified facts
   • Provides contrast for calibration

Generator Classes

Class	Purpose	File
AntiHallucinationGenerator	Generate refusal examples	src/realign/data/antihallucination_generator.py
RefusalPreferenceGenerator	Preference pairs for refusal	src/realign/data/refusal_preference_generator.py
CalibrationTrainer	Confidence calibration	src/realign/backends/olmo/calibration_trainer.py

Confidence Levels

CONFIDENCE_LEVELS = {
    "high": "I am confident that...",
    "medium": "I believe that...",
    "low": "I'm not certain, but...",
    "uncertain": "I don't know..."
}

Training Data Mix

Recommended anti-hallucination mix:

• 40% - Appropriate refusals (unknown facts)
• 30% - Hedged uncertainty (partial knowledge)
• 20% - Confident answers (verified facts)
• 10% - Edge cases (ambiguous queries)

Commands

# Generate anti-hallucination data
python -m realign.data.antihallucination_generator \
  --input sft.jsonl \
  --output antihall.jsonl \
  --refusal-ratio 0.4 \
  --uncertainty-ratio 0.3

# Generate calibration data
python scripts/calibration_generator.py \
  --count 5000 \
  --output calibration.jsonl

Evaluation

# Check TruthfulQA score (target: ≥60%)
realign evaluate --model your-model --benchmark truthfulqa

# Check calibration ECE (target: ≤0.10)
realign evaluate --model your-model --benchmark calibration

Acceptance Criteria

• All generator classes documented
• Confidence levels defined
• Training data mix ratios
• CLI commands provided
• Evaluation benchmarks documented

Related

• Issue feat(skills): Update data-curation-workflow skill with A-grade 9-stage pipeline #305 (data-curation-workflow)
• Issue feat(skills): Create training-methods skill documenting 8 stable training methods #306 (training-methods)
• Issue feat(skills): Create dpo-rlvr-generation skill for preference and verification data #307 (dpo-rlvr-generation)

[akaszubski]

Implemented in PR #324 (batch-20260205-233714)