Kamoo re Koetlisitseng Shannon V1.5 ho Nahana ho Sebelisa GRPO

1. Khothatso: Hobaneng Mehlala ea ho Nahana?

Bakeng sasehlopha sa AI red teamlipatlisiso, ho utloisisakamooAI e fihlella liphetho tse ka bang kotsi ho bohlokoa joalo ka ho ithuta liphetho ka botsona. Mehlala ea ho nahana ea Shannon V1.5 e senola ketane e felletseng ea menahano, e nolofalletsa bafuputsi ho:

• Latela tsela ea ho nahana e lebisang holiphello tsa AI tse sa laolehengboitšoaro
• Hlokomela moo li-guardrail li neng li tla kena lipakeng teng ka tloaelo (e bontšangbohlokoa ba li-guardrail tsa AI)
• Ntshetsa pele mekhoa e matla haholoanyane ea polokeho e lebisang mekhoeng e itseng ea ho nahana
• Theha linyeoe tse betere tsa liteko tse hanyetsang ho ipapisitsoe le liketane tsa menahano tse hlokometsoeng

2. Ho Utloisisa GRPO

Group Relative Policy Optimization (GRPO)ke tsoelo-pele ho feta mekhoa ea setso ea RLHF e nolofalletsang koetliso e tsitsitseng le e sebetsang hantle ea bokhoni ba ho nahana. E entsoe ke DeepSeek AI, e ipakile e sebetsa hantle haholo bakeng sa koetliso ea ketane ea menahano.

Hobaneng GRPO ho Feta RLHF ea Setso?

Motheo oa Lipalo oa GRPO

GRPO e ntlafatsa leano ka ho bapisa likarabo ka har'a lihlopha ho fapana le khahlanong le mohlala oa moputso o felletseng:

Papiso ena e amanang e na le melemo e mengata:

• Ho Tloaeleha:E ikamahanya ka bo eona le bothata bo fapaneng ho pholletsa le litaelo
• Botsitso:E fokotsa phapang ho likhakanyo tsa gradient
• Bokgoni:Ha ho hlokahale mohlala o arohaneng oa moputso

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 Distillation

Ho qala bokhoni ba ho nahana ba Shannon V1.5, re ile ra ntša mekhoa ea ketane ea menahano ho tsoa mehlaleng ea ho nahana ea DeepSeek. Sena se fane ka litsela tsa CoT tsa boleng bo holimo ho koetlisa hlooho ea rona ea ho nahana.

Sebopeho sa Dataset sa DeepSeek

Mokhoa oa ho Bokella Litsela

Re bokeletse litsela tsa ho nahana libakeng tse fapaneng ho netefatsa hore ho nahana ho akaretsa hohle:

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. Sebopeho sa Hlooho ea Monahano

Mehlala ea Shannon V1.5 e kenyelletsa e inehetsenghlooho ea monahanoe hlahisang litsela tsa ho nahana tse hlakileng pele ho sephetho sa ho qetela. Keketso ena ea meralo e nolofalletsa CoT e bonaletsang ntle le ho fetola sebopeho sa motheo sa Mixtral.

Ts'ebetso ea Hlooho ea Monahano

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. Mokhoa oa Koetliso

Mohato oa 1: Koetliso ea Pele ea Hlooho ea Monahano

Pele, re koetlisa hlooho ea monahano pele ho litsela tsa CoT tse tsoang ho DeepSeek re sebelisa tahlehelo e tloaelehileng ea cross-entropy:

# 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

Mohato oa 2: Tokiso e Ntle ea GRPO

Kamora koetliso ea pele, re sebelisa GRPO ho ntlafatsa boleng ba monahano re sebelisa lipapiso tse amanang le sehlopha:

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()
        }

Mohato oa 3: Boikhethelo ba Sehlopha se Sefubelu

Qetellong, re lokisa hape maemong a boipelaetso ho netefatsa hore litsela tsa monahano li senola hantle mabaka aAI e sa hlahlobjoang e latelangtlhahlobo:

6. Liphetho le Tlhahlobo

Litekanyo tsa Boleng ba Monahano

Lintlha tsa Bohlokoa

• Bonaletsi bo ntlafetse haholo:Ho tloha ho 12% ho ea ho 97.8% ea monahano joale e hlalosoa ka ho hlaka
• Nepahalo ea monahano e eketsehile:Monahano o hlakileng o ntlafalitse boleng ba karabo ea ho qetela ka lintlha tse 12+
• Boleng ba sehlopha se sefubelu bo netefalitsoe:Bafuputsi ba tšireletso ba tlaleha hore litsela tsa monahano ke "tsa bohlokoa" bakeng sa ho utloisisa mabaka a ho sebelisa hampe
• GRPO e fetile RLHF:Lintlha tsa kopano tse ntle ka 15% ha li bapisoa le mokhoa oa setso