self_evaluation.py

#!/usr/bin/env python3
"""
AETHER-Micro Self-Evaluation Head

4-dimensional quality assessment for RLP training
"""

import torch
import torch.nn as nn

from .configuration_aether_micro import AETHERMicroConfig


class AETHERMicroSelfEvalHead(nn.Module):
    """
    Self-Evaluation Head for Quality Assessment

    Features:
    - 4-dim Quality: factuality, coherence, completeness, specificity
    - Overall Score: 1-dim sigmoid for aggregated quality
    - Parameter Count: ~600K

    Architecture:
        hidden_size → 1024 → 512 → 4 quality dims
                            → 512 → 1 overall score

    Quality Dimensions:
        - factuality: 사실성/정확성 (0~1)
        - coherence: 일관성/논리성 (0~1)
        - completeness: 완결성/충분성 (0~1)
        - specificity: 구체성/상세성 (0~1)

    Overall Score:
        - Aggregated quality (0~1)
        - Used for RLP reward signal
    """

    def __init__(self, config: AETHERMicroConfig):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.self_eval_dims = config.self_eval_dims

        # Shared encoder
        self.encoder = nn.Sequential(
            nn.Linear(self.hidden_size, 1024),
            nn.GELU(),
            nn.Linear(1024, 512),
            nn.GELU()
        )

        # Quality dimensions (4-dim)
        self.quality_head = nn.Linear(512, self.self_eval_dims)

        # Overall score (1-dim)
        self.overall_head = nn.Linear(512, 1)

    def forward(self, hidden_states: torch.Tensor) -> tuple:
        """
        Args:
            hidden_states: (batch_size, seq_length, hidden_size)

        Returns:
            quality: (batch_size, seq_length, 4) - [factuality, coherence, completeness, specificity]
            overall: (batch_size, seq_length, 1) - aggregated quality score
        """
        # Encode
        encoded = self.encoder(hidden_states)  # (B, S, 512)

        # Quality dimensions (sigmoid)
        quality = torch.sigmoid(self.quality_head(encoded))  # (B, S, 4)

        # Overall score (sigmoid)
        overall = torch.sigmoid(self.overall_head(encoded))  # (B, S, 1)

        return quality, overall


# ========================================
# Parameter Count Verification
# ========================================

def count_self_eval_parameters(config: AETHERMicroConfig) -> int:
    """
    Calculate Self-Evaluation Head parameter count

    Expected: ~600K parameters

    Breakdown:
        hidden_size → 1024: 896 × 1024 = 917,504
        1024 → 512: 1024 × 512 = 524,288
        512 → 4 (quality): 512 × 4 = 2,048
        512 → 1 (overall): 512 × 1 = 512
        Total: 917,504 + 524,288 + 2,048 + 512 = 1,444,352

    Note: 기존 NIPA 모델과 약간 차이 (1.4M vs 600K)
          실제 기존 모델 구조 재확인 필요
    """
    head = AETHERMicroSelfEvalHead(config)
    return sum(p.numel() for p in head.parameters())


if __name__ == "__main__":
    from .configuration_aether_micro import AETHERMicroConfig

    config = AETHERMicroConfig()
    param_count = count_self_eval_parameters(config)
    print(f"Self-Evaluation Head Parameters: {param_count:,}")