Kako smo trenirali Shannon V1.5 za razmišljanje z uporabo GRPO

1. Motivacija: Zakaj misleči modeli?

Zardečo ekipo umetne inteligenceraziskave, je razumevanjekakoumetna inteligenca pride do potencialno škodljivih rezultatov, enako pomembno kot preučevanje samih rezultatov. Misleči modeli Shannon V1.5 razkrivajo celotno verigo misli, kar raziskovalcem omogoča:

• Sledenje poti sklepanja, ki vodi donecenzuriranih posledičnihvedenj umetne inteligence
• Ugotovitev, kje bi običajno posredovale varnostne ograje (kar kaže napomen varnostnih ograj umetne inteligence)
• Razvoj robustnejših varnostnih mehanizmov, ki ciljajo na specifične vzorce sklepanja
• Ustvarjanje boljših nasprotnih testnih primerov na podlagi opazovanih verig sklepanja

2. Razumevanje GRPO

Optimizacija politike relativne skupine (GRPO)je napredek v primerjavi s tradicionalnimi metodami RLHF, ki omogoča stabilnejše in učinkovitejše usposabljanje sposobnosti sklepanja. Razvit s strani DeepSeek AI, se je izkazal za še posebej učinkovitega pri usposabljanju verige misli.

Zakaj GRPO namesto tradicionalnega RLHF?

Matematična osnova GRPO

GRPO optimizira politiko s primerjanjem odgovorov znotraj skupin namesto proti absolutnemu modelu nagrajevanja:

Ta relativna primerjava ima več prednosti:

• Normalizacija:Samodejno se prilagaja različnim težavnostim med pozivi
• Stabilnost:Zmanjšuje varianco v ocenah gradienta
• Učinkovitost:Ni potreben ločen model nagrajevanja

def compute_grpo_loss(
    policy_logprobs: torch.Tensor,
    rewards: torch.Tensor,
    group_size: int = 8
) -> torch.Tensor:
    """
    Compute GRPO loss with group-relative reward normalization.
    
    Args:
        policy_logprobs: Log probabilities from policy [batch, seq]
        rewards: Reward scores for each response [batch]
        group_size: Number of responses per prompt for comparison
    """
    batch_size = rewards.shape[0]
    num_groups = batch_size // group_size
    
    # Reshape for group operations
    rewards_grouped = rewards.view(num_groups, group_size)
    logprobs_grouped = policy_logprobs.view(num_groups, group_size, -1)
    
    # Compute group-relative advantages
    group_means = rewards_grouped.mean(dim=1, keepdim=True)
    group_stds = rewards_grouped.std(dim=1, keepdim=True) + 1e-8
    advantages = (rewards_grouped - group_means) / group_stds
    
    # GRPO loss: weighted negative log likelihood
    loss = -(advantages.unsqueeze(-1) * logprobs_grouped).sum(dim=-1).mean()
    
    return loss

3. Destilacija DeepSeek

Za zagon miselnih sposobnosti Shannon V1.5 smo destilirali vzorce verige misli iz DeepSeekovih modelov sklepanja. To je zagotovilo visokokakovostne sledi CoT za usposabljanje naše misleče glave.

Sestava nabora podatkov DeepSeek

Postopek zbiranja sledi

Zbrali smo sledi razmišljanja iz različnih domen, da zagotovimo celovito pokritost sklepanja:

class DeepSeekDistiller:
    """Distill chain-of-thought traces from DeepSeek models."""
    
    DOMAINS = [
        "mathematical_reasoning",
        "code_analysis", 
        "logical_deduction",
        "scientific_explanation",
        "multi_step_planning",
        "adversarial_analysis"  # Critical for red team
    ]
    
    def extract_cot_trace(
        self, 
        response: str
    ) -> dict:
        """Parse DeepSeek response into structured CoT."""
        
        # DeepSeek uses ... tags
        think_match = re.search(
            r'(.*?)', 
            response, 
            re.DOTALL
        )
        
        if not think_match:
            return None
            
        thinking = think_match.group(1)
        final_answer = response.split('')[-1].strip()
        
        # Parse individual reasoning steps
        steps = self.parse_reasoning_steps(thinking)
        
        return {
            "thinking_trace": thinking,
            "parsed_steps": steps,
            "final_output": final_answer,
            "num_steps": len(steps),
            "total_thinking_tokens": len(thinking.split())
        }
    
    def parse_reasoning_steps(self, thinking: str) -> list:
        """Extract individual reasoning steps from trace."""
        # Split on common step indicators
        step_patterns = [
            r'\n\d+\.',           # "1. ", "2. "
            r'\nStep \d+:',       # "Step 1:"
            r'\n(?:First|Next|Then|Finally),',
            r'\n- '              # Bullet points
        ]
        
        combined_pattern = '|'.join(step_patterns)
        steps = re.split(combined_pattern, thinking)
        
        return [s.strip() for s in steps if s.strip()]

4. Arhitektura miselne glave

Modeli Shannon V1.5 vključujejo namenskomiselno glavoki generira eksplicitne sledi sklepanja pred končnim izhodom. Ta arhitekturni dodatek omogoča pregleden CoT brez spreminjanja osnovne arhitekture Mixtral.

Implementacija miselne glave

class ThinkingHead(nn.Module):
    """
    Dedicated thinking module for Shannon V1.5.
    Generates explicit chain-of-thought traces.
    """
    
    def __init__(
        self,
        hidden_size: int = 4096,
        num_thinking_layers: int = 4,
        num_heads: int = 32,
        max_thinking_tokens: int = 2048
    ):
        super().__init__()
        
        self.hidden_size = hidden_size
        self.max_thinking_tokens = max_thinking_tokens
        
        # Special tokens
        self.think_start = nn.Parameter(torch.randn(1, 1, hidden_size))
        self.think_end = nn.Parameter(torch.randn(1, 1, hidden_size))
        
        # Thinking transformer layers
        self.thinking_layers = nn.ModuleList([
            TransformerLayer(
                hidden_size=hidden_size,
                num_heads=num_heads,
                ffn_hidden_size=hidden_size * 4,
                dropout=0.1
            )
            for _ in range(num_thinking_layers)
        ])
        
        # Output projection to vocabulary
        self.output_proj = nn.Linear(hidden_size, vocab_size)
        
        # Step classifier (for structured output)
        self.step_classifier = nn.Linear(hidden_size, 5)  # 5 step types
    
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        generate_steps: bool = True
    ) -> dict:
        """
        Generate thinking trace from input hidden states.
        
        Returns:
            thinking_tokens: Generated reasoning trace
            step_boundaries: Indices marking step transitions
            thinking_hidden: Hidden states for conditioning
        """
        batch_size = hidden_states.shape[0]
        
        # Prepend thinking start token
        thinking_input = torch.cat([
            self.think_start.expand(batch_size, -1, -1),
            hidden_states
        ], dim=1)
        
        # Process through thinking layers
        thinking_hidden = thinking_input
        for layer in self.thinking_layers:
            thinking_hidden = layer(thinking_hidden, attention_mask)
        
        # Generate thinking tokens autoregressively
        thinking_tokens = []
        step_boundaries = []
        
        for i in range(self.max_thinking_tokens):
            logits = self.output_proj(thinking_hidden[:, -1, :])
            next_token = logits.argmax(dim=-1)
            
            # Check for step boundaries
            step_type = self.step_classifier(thinking_hidden[:, -1, :])
            if step_type.argmax(dim=-1) != 0:  # 0 = continue
                step_boundaries.append(i)
            
            thinking_tokens.append(next_token)
            
            # Check for think_end
            if next_token == self.think_end_token_id:
                break
            
            # Update for next iteration
            # ... (autoregressive generation logic)
        
        return {
            "thinking_tokens": torch.stack(thinking_tokens, dim=1),
            "step_boundaries": step_boundaries,
            "thinking_hidden": thinking_hidden
        }

5. Cevovod usposabljanja

Faza 1: Predusposabljanje miselne glave

Najprej predusposobimo miselno glavo na sledi CoT, destilirane iz DeepSeeka, z uporabo standardne izgube navzkrižne entropije:

# Thinking Head Pre-training Configuration
model:
  base: shannon-ai/v1-deep  # Start from GPT-5 distilled model
  thinking_head:
    num_layers: 4
    hidden_size: 4096
    max_tokens: 2048

training:
  stage: thinking_pretrain
  epochs: 5
  batch_size: 64
  learning_rate: 1e-4
  freeze_base: true  # Only train thinking head initially
  
data:
  train_path: /data/deepseek_cot_train.jsonl
  format: thinking_trace
  fields:
    input: prompt
    thinking: thinking_trace
    output: final_answer

Faza 2: Fino uglaševanje GRPO

Po predusposabljanju uporabimo GRPO za izboljšanje kakovosti razmišljanja z uporabo skupinsko-relativnih primerjav:

class GRPOTrainer:
    """GRPO trainer for thinking model optimization."""
    
    def __init__(
        self,
        model: ThinkingModel,
        group_size: int = 8,
        kl_coef: float = 0.1
    ):
        self.model = model
        self.group_size = group_size
        self.kl_coef = kl_coef
        self.ref_model = copy.deepcopy(model)
        self.ref_model.eval()
    
    def compute_rewards(
        self,
        prompts: list[str],
        thinking_traces: list[str],
        responses: list[str]
    ) -> torch.Tensor:
        """
        Compute rewards for thinking quality.
        Multiple signals combined for comprehensive evaluation.
        """
        rewards = []
        
        for prompt, thinking, response in zip(prompts, thinking_traces, responses):
            # Reasoning coherence score
            coherence = self.evaluate_coherence(thinking)
            
            # Step structure quality
            structure = self.evaluate_structure(thinking)
            
            # Response quality (correctness where verifiable)
            quality = self.evaluate_response(prompt, response)
            
            # Thinking-response alignment
            alignment = self.evaluate_alignment(thinking, response)
            
            # Combined reward
            reward = (
                0.3 * coherence +
                0.2 * structure +
                0.3 * quality +
                0.2 * alignment
            )
            rewards.append(reward)
        
        return torch.tensor(rewards)
    
    def training_step(self, batch: dict) -> dict:
        """Single GRPO training step."""
        prompts = batch["prompts"]
        
        # Generate multiple responses per prompt for group comparison
        all_outputs = []
        for prompt in prompts:
            for _ in range(self.group_size):
                output = self.model.generate_with_thinking(
                    prompt,
                    temperature=0.8,  # Diversity for comparison
                    do_sample=True
                )
                all_outputs.append(output)
        
        # Compute rewards
        rewards = self.compute_rewards(
            prompts=[p for p in prompts for _ in range(self.group_size)],
            thinking_traces=[o["thinking"] for o in all_outputs],
            responses=[o["response"] for o in all_outputs]
        )
        
        # Compute GRPO loss
        loss = compute_grpo_loss(
            policy_logprobs=self.get_logprobs(all_outputs),
            rewards=rewards,
            group_size=self.group_size
        )
        
        # Add KL penalty against reference model
        kl_div = self.compute_kl_divergence(all_outputs)
        total_loss = loss + self.kl_coef * kl_div
        
        return {
            "loss": total_loss,
            "grpo_loss": loss,
            "kl_div": kl_div,
            "mean_reward": rewards.mean()
        }

Faza 3: Specializacija rdeče ekipe

Končno, nadalje uglašujemo na nasprotne scenarije, da zagotovimo, da sledi razmišljanja pravilno razkrivajo sklepanje zanecenzurirano AI posledičnoanalizo:

6. Rezultati in analiza

Metrike kakovosti razmišljanja

Ključne ugotovitve

• Preglednost se je dramatično izboljšala:Od 12% do 97.8% sklepanja je zdaj eksplicitno verbaliziranega
• Natančnost sklepanja se je povečala:Eksplicitno razmišljanje je izboljšalo kakovost končnega odgovora za 12+ točk
• Vrednost rdeče ekipe potrjena:Varnostni raziskovalci poročajo, da so sledi razmišljanja »neprecenljive« za razumevanje sklepanja izkoriščanja
• GRPO je presegel RLHF:15% boljše ocene koherence v primerjavi s tradicionalnim pristopom