Biz Shannon V1.5 ni GRPO yordamida fikrlashga qanday o'rgatdik

1. Motivatsiya: Nima uchun fikrlash modellari?

UchunAI qizil jamoasitadqiqotida, tushunishqandayAI ning potentsial zararli natijalarga qanday erishishini tushunish, natijalarning o'zini o'rganish kabi muhimdir. Shannon V1.5 ning fikrlash modellari fikrlash zanjirini to'liq ochib beradi, bu tadqiqotchilarga quyidagilarni amalga oshirish imkonini beradi:

• Sabab-natija yo'lini kuzatishtsenzurasiz AI natijasida yuzaga keladiganxatti-harakatlar
• Himoya mexanizmlari odatda qayerda aralashishini aniqlash (ko'rsatib berish)AI himoya mexanizmlarining ahamiyati)
• Muayyan fikrlash shakllariga qaratilgan yanada mustahkam xavfsizlik mexanizmlarini ishlab chiqish
• Kuzatilgan fikrlash zanjirlariga asoslangan holda yaxshiroq raqib sinov holatlarini yaratish

2. GRPO ni tushunish

Group Relative Policy Optimization (GRPO)an'anaviy RLHF usullaridan ustun bo'lgan yutuq bo'lib, fikrlash qobiliyatlarini yanada barqaror va samarali o'qitish imkonini beradi. DeepSeek AI tomonidan ishlab chiqilgan bo'lib, u fikrlash zanjirini o'qitish uchun ayniqsa samarali ekanligini isbotladi.

Nima uchun GRPO an'anaviy RLHF dan ustun?

GRPO Matematik Asosi

GRPO siyosatni mutlaq mukofot modeliga qarshi emas, balki guruhlar ichidagi javoblarni taqqoslash orqali optimallashtiradi:

Bu nisbiy taqqoslash bir qancha afzalliklarga ega:

• Normalizatsiya:So'rovlar bo'yicha turli qiyinchiliklarga avtomatik ravishda moslashadi
• Barqarorlik:Gradient baholaridagi dispersiyani kamaytiradi
• Samaradorlik:Alohida mukofot modeli kerak emas

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. DeepSeek Distillatsiyasi

Shannon V1.5 ning fikrlash qobiliyatlarini boshlash uchun biz DeepSeek ning fikrlash modellaridan fikrlash zanjiri (CoT) shakllarini distillatsiya qildik. Bu bizning fikrlash boshimizni o'qitish uchun yuqori sifatli CoT izlarini taqdim etdi.

DeepSeek ma'lumotlar to'plami tarkibi

Izlarni yig'ish jarayoni

Biz keng qamrovli mantiqiy fikrlashni ta'minlash uchun turli sohalardagi fikrlash izlarini to'pladik:

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. Fikrlash boshining arxitekturasi

Shannon V1.5 modellari maxsusfikrlash boshiniyakuniy natijadan oldin aniq mantiqiy fikrlash izlarini yaratadi. Bu arxitekturaviy qo'shimcha asosiy Mixtral arxitekturasini o'zgartirmasdan shaffof CoTni ta'minlaydi.

Fikrlash boshini amalga oshirish

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. O'qitish jarayoni

1-bosqich: Fikrlash boshini oldindan o'qitish

Birinchidan, biz fikrlash boshini DeepSeek-distillangan CoT izlari bo'yicha standart kross-entropiya yo'qotishidan foydalanib oldindan o'qitamiz:

# 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

2-bosqich: GRPO nozik sozlash

Oldindan o'qitishdan so'ng, biz GRPO'ni guruhga nisbatan taqqoslashlardan foydalanib fikrlash sifatini yaxshilash uchun qo'llaymiz:

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()
        }

3-bosqich: Qizil jamoa ixtisoslashuvi

Nihoyat, biz fikrlash izlari mantiqiy fikrlashni to'g'ri ochib berishini ta'minlash uchun raqib stsenariylarida qo'shimcha sozlashlar qilamizsenzurasiz AI natijaviytahlil uchun:

6. Natijalar va tahlil

Fikrlash sifati ko'rsatkichlari

Asosiy topilmalar

• Shaffoflik sezilarli darajada yaxshilandi:Mantiqiy fikrlashning 12% dan 97.8% gacha qismi endi aniq og'zaki ifodalangan
• Mantiqiy fikrlash aniqligi oshdi:Aniq fikrlash yakuniy javob sifatini 12+ ballga oshirdi
• Qizil jamoa qiymati tasdiqlandi:Xavfsizlik tadqiqotchilari fikrlash izlari ekspluatatsiya mantiqini tushunish uchun "bepul" ekanligini ta'kidlashadi
• GRPO RLHF'dan ustun chiqdi:An'anaviy yondashuvga nisbatan 15% yaxshiroq izchillik ballari