quality-scoring（质量评分）

Quality Scoring

Multi-dimensional assessment for training data quality.

When Activates

Quality assessment, data scoring, multi-dimensional evaluation, IFD scoring, factuality checks, reasoning validation, training data prep

Core Concepts

Quality Scorers (6 Types)

Fast to comprehensive scoring approaches:

1. FastIFD - Instruction-following difficulty (10-20x faster)
2. Quality - LLM-based quality (Qwen3-30B, 0.85 ex/s)
3. MultiDimensional - 5-dimension composite
4. LLMQuality - Multi-backend (MLX/OpenRouter)
5. Ensemble - Cross-model ensemble
6. Tulu3 - Multi-dimensional reference (training_metrics.py)

Quality Dimensions (6 Metrics)

1. IFD Score (0.0-1.0) - Instruction-following difficulty
2. Factuality (0.0-1.0) - Hallucination detection
3. Reasoning (0.0-1.0) - Step-by-step logic quality
4. Diversity (0.0-1.0) - Dataset-level diversity
5. Domain (0.0-1.0) - Domain-specific relevance
6. LLM Quality (1-10) - Tulu3 comprehensive score

Training Thresholds

Quick Reference

IFD Score Calculation

from training_metrics import calculate_ifd_score

# IFD = PPL(response) / PPL(response|instruction)
ifd_score = calculate_ifd_score(
    instruction="Explain quantum computing",
    response="Quantum computing uses qubits..."
)
# Higher score = more challenging

DPO Pair Validation

from training_metrics import validate_dpo_pairs

# Validate chosen/rejected quality gap
is_valid = validate_dpo_pairs(
    chosen_score=9.2,  # High quality
    rejected_score=5.8  # Low quality
)
# Ensures quality gap ≥0.15

REQUIRED: DPO Multi-Dimensional Scoring

Every DPO pair MUST have multi-dimensional quality scores before training.

This is a hard requirement — DPO data without quality scores will learn shortcuts (e.g., "longer = better") instead of genuine preference signal.

Required output fields per pair:

• chosen_score (float): Composite quality score for chosen response
• rejected_score (float): Composite quality score for rejected response
• margin (float): chosen_score - rejected_score (must be ≥3.0)

Length bias audit (MUST run before DPO training):

from pathlib import Path
from training_metrics import validate_dpo_pairs

metrics = validate_dpo_pairs(dpo_path=Path("dpo_pairs.jsonl"))

# Check length bias
longer_chosen = sum(1 for p in metrics.pairs if len(p.chosen) > len(p.rejected))
length_bias = longer_chosen / metrics.total_pairs

if length_bias > 0.70:
    raise ValueError(
        f"DPO length bias {length_bias:.0%} > 70% threshold.\n"
        f"Model will learn 'longer = better' shortcut.\n"
        f"Fix: Score by quality dimensions, not length."
    )

# Check quality scores present
missing = sum(1 for p in metrics.pairs if p.chosen_score is None)
if missing > 0:
    raise ValueError(f"{missing} pairs missing quality scores — run scoring first")

Scoring workflow:

1. Generate DPO pairs (dpo-rlvr-generation skill)
2. Score all pairs with multi-dimensional scorer (this skill)
3. Filter by quality margin ≥3.0
4. Audit length bias ≤70%
5. Only then proceed to training

RLVR Verifiability

from training_metrics import assess_rlvr_verifiability

# Assess reasoning trace verifiability
verifiable = assess_rlvr_verifiability(
    reasoning_trace="Step 1: ...\nStep 2: ...",
    domain="math"
)
# Math/coding: 90%+ verifiable required

Progressive Disclosure

Detailed guides: See docs/*.md

• docs/quality-scorers.md - 6 scorer implementations
• docs/quality-dimensions.md - 6 dimension definitions
• docs/training-thresholds.md - Thresholds, CLI, distributed performance

Security Considerations

Input Validation (CWE-20)

• Validate score ranges (0.0-1.0 or 1-10)
• Sanitize data inputs before scoring
• Check threshold values before application

Path Traversal (CWE-22)

• Sanitize file paths for data loading
• Whitelist directories for training data
• Validate output paths for scored datasets

Security Patterns (training_metrics.py)

from pathlib import Path

def safe_load_data(data_path: str) -> dict:
    """Load data with path validation."""
    # Validate path within allowed directory
    path = Path(data_path).resolve()
    if not str(path).startswith('/allowed/data/'):
        raise ValueError(f"Path outside allowed directory: {path}")

    # Load safely
    return json.loads(path.read_text())

Distributed Performance

Single Machine Performance

• M4 Max: ~0.85 ex/s (Qwen3-30B)
• M3 Ultra: ~0.85 ex/s (Qwen3-30B)

Parallel Processing

• Combined throughput: ~1.7 ex/s (50/50 split)
• Scaling: Linear with machine count
• Bottleneck: Model inference, not I/O

CLI Commands

# Score dataset with FastIFD
python -m training_metrics score \
  --input data/train.jsonl \
  --output data/scored.jsonl \
  --scorer fastifd \
  --threshold 0.3

# Multi-dimensional scoring
python -m training_metrics score \
  --input data/train.jsonl \
  --output data/scored.jsonl \
  --scorer multidim \
  --quality-threshold 8.0 \
  --ifd-threshold 0.5

# DPO pair filtering
python -m training_metrics filter_dpo \
  --input data/dpo_pairs.jsonl \
  --output data/filtered_pairs.jsonl \
  --chosen-threshold 9.0 \
  --rejected-threshold 6.0

# RLVR verifiability check
python -m training_metrics assess_rlvr \
  --input data/rlvr_traces.jsonl \
  --output data/verified.jsonl \
  --domain math \
  --threshold 0.9

Related Skills

• data-distillation - IFD methodology and KenLM filtering
• preference-data-quality - DPO and RLVR metrics
• python-standards - Code quality standards

Library Integration

Primary library: training_metrics.py

Key functions:

• calculate_ifd_score() - IFD calculation
• validate_dpo_pairs() - DPO pair validation
• assess_rlvr_verifiability() - RLVR assessment
• score_quality() - Multi-dimensional scoring
• ensemble_score() - Cross-model ensemble

Key Takeaways

1. 6 scorers - FastIFD (fast) to Ensemble (comprehensive)
2. 6 dimensions - IFD, Factuality, Reasoning, Diversity, Domain, LLM Quality
3. Training thresholds - SFT ≥8.0, DPO chosen ≥9.0, RLVR ≥9.0
4. IFD score - PPL(response) / PPL(response|instruction), higher = harder
5. Security - CWE-20 (input validation), CWE-22 (path traversal)
6. Distributed - ~1.7 ex/s with 2 machines (linear scaling)
7. CLI commands - training_metrics module for all operations
8. Integration - Use training_metrics library functions
9. DPO pairs - Chosen ≥9.0, Rejected ≤6.0, gap ≥0.15
10. RLVR - Math/coding 90%+ verifiable, general 80%+
11. DPO scoring REQUIRED - Every pair must have chosen_score, rejected_score, margin before training
12. Length bias audit - ≤70% of pairs where chosen is longer (prevents "longer = better" shortcut)