Yadda Muka Horar da Shannon V1.5 don Yin Tunani ta Amfani da GRPO

1. Dalili: Me yasa Samfuran Tunani?

DonAI red teambincike, fahimtaryaddawani AI ke kaiwa ga abubuwan da ka iya zama masu cutarwa yana da mahimmanci kamar nazarin abubuwan da kansu. Samfuran tunani na Shannon V1.5 suna bayyana cikakken jerin-tunani, yana bawa masu bincike damar:

• Gano hanyar tunani da ke kaiwa gasakamakon AI mara tantancewahalaye
• Gano inda matakan tsaro za su shiga tsakani a al'ada (yana nunamahimmancin matakan tsaro na AI)
• Haɓaka ingantattun hanyoyin tsaro da ke nufin takamaiman tsarin tunani
• Ƙirƙiri ingantattun shari'o'in gwaji na adawa bisa ga jerin tunani da aka lura

2. Fahimtar GRPO

Inganta Manufofin Dangantaka na Rukuni (GRPO)ci gaba ne akan hanyoyin RLHF na gargajiya wanda ke ba da damar horar da iyawar tunani mai tsayayye da inganci. An haɓaka shi ta DeepSeek AI, ya tabbatar da cewa yana da tasiri musamman don horar da jerin-tunani.

Me yasa GRPO ya fi RLHF na Gargajiya?

Tushen Lissafi na GRPO

GRPO yana inganta manufofi ta hanyar kwatanta amsoshi a cikin rukuni maimakon a kan cikakken samfurin lada:

Wannan kwatancen dangantaka yana da fa'idodi da yawa:

• Daidaitawa:Yana daidaitawa ta atomatik don bambancin wahala a cikin umarni
• Tsayayye:Yana rage bambancin a cikin ƙididdigar gradient
• Inganci:Ba a buƙatar samfurin lada daban

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. Tsarkake DeepSeek

Don fara iyawar tunani na Shannon V1.5, mun tsarkake tsarin jerin-tunani daga samfuran tunani na DeepSeek. Wannan ya samar da ingantattun bibiyoyin CoT don horar da kan tunaninmu.

Tsarin Bayanan DeepSeek

Tsarin Tattara Alamomi

Mun tattara alamomin tunani daga fannoni daban-daban don tabbatar da cikakken tunani:

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. Tsarin Kan Tunani

Samfuran Shannon V1.5 sun haɗa da keɓaɓɓenkan tunaniwanda ke samar da alamomin tunani a fili kafin fitarwa ta ƙarshe. Wannan ƙarin tsarin yana ba da damar CoT mai bayyananniya ba tare da canza ainihin tsarin Mixtral ba.

Aiwatar da Kan Tunani

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. Tsarin Horarwa

Mataki na 1: Horar da Kan Tunani na Farko

Da farko, mun horar da kan tunani a kan alamomin CoT da aka tace daga DeepSeek ta amfani da asarar cross-entropy na yau da kullun:

# 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

Mataki na 2: Gyaran GRPO

Bayan horarwa ta farko, mun yi amfani da GRPO don inganta ingancin tunani ta amfani da kwatancen da suka shafi rukuni:

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()
        }

Mataki na 3: Kwarewar Ƙungiyar Ja

A ƙarshe, mun ƙara gyara kan yanayin adawa don tabbatar da cewa alamomin tunani sun bayyana tunani yadda ya kamata donAI mara tantancewa sakamakonbincike:

6. Sakamako & Bincike

Ma'aunin Ingancin Tunani

Mahimman Abubuwan Da Aka Gano

• Bayyananniya ta inganta sosai:Daga 12% zuwa 97.8% na tunani yanzu an bayyana shi a fili
• Daidaiton tunani ya karu:Tunanin da aka bayyana a fili ya inganta ingancin amsa ta ƙarshe da maki 12+
• An tabbatar da darajar ƙungiyar ja:Masu binciken tsaro sun ba da rahoton cewa alamomin tunani suna da "daraja" don fahimtar tunanin cin zarafi
• GRPO ya fi RLHF aiki:Maki 15% mafi kyawun haɗin kai idan aka kwatanta da tsarin gargajiya