Data Quality Filtering — RLHF, DPO, and PPO

Training data quality is the primary determinant of aligned model behavior. Even sophisticated alignment techniques like RLHF fail when applied to noisy, inconsistent, or adversarial examples.

Preference Scoring Heuristics

The simplest filtering approach assigns quality scores to training pairs based on heuristic signals:

def compute_quality_score(prompt, response, metadata):
    score = 0.0
    
    # Length heuristics
    if len(response) < 50:
        score -= 0.5  # Likely too short
    if len(response) > 4000:
        score -= 0.3  # Possibly rambling
    
    # Repetition penalties
    ngrams = extract_ngrams(response, n=3)
    repetition_ratio = max(counts.values()) / sum(counts.values())
    if repetition_ratio > 0.3:
        score -= 0.4
    
    # Coherence signals (from metadata)
    if metadata.get("thumbs_up", 0) > metadata.get("thumbs_down", 0):
        score += 0.3
    
    # Specificity rewards
    if has_code_block(response):
        score += 0.2
    if has_citations(response):
        score += 0.15
    
    return score

Automated Quality Detection

Modern pipelines use classifiers to identify high-quality responses:

# Train a quality classifier on human-rated samples
python train_quality_classifier.py \
    --training-data human_ratings.parquet \
    --model bert-base \
    --output quality_classifier.pt

# Filter dataset with threshold
python filter_dataset.py \
    --input raw_preferences.jsonl \
    --classifier quality_classifier.pt \
    --threshold 0.75 \
    --output filtered_preferences.jsonl

Pairs vs. Singles Filtering

Filtering individual responses differs from filtering preference pairs:

Aspect	Single Response	Preference Pair
Criterion	Response quality	Relative quality
Signal	Classifier score	Consistency of preference
Failure mode	Misses good responses	Requires both responses valid

Adversarial Data Handling

Attackers intentionally create preference pairs designed to corrupt training:

def detect_adversarial_pair(prompt, chosen, rejected):
    # Check for prompt injection patterns
    if contains_injection_pattern(prompt):
        return True
    
    # Check for suspiciously close responses (hard to distinguish)
    similarity = compute_embedding_similarity(chosen, rejected)
    if similarity > 0.95:
        return True  # Both responses nearly identical
    
    # Check for reversed quality (jailbreak attempts)
    if is_jailbreak_prompt(prompt):
        if "sorry" in chosen.lower():
            return False  # Good response to jailbreak
        else:
            return True  # Suspicious
    
    return False