Ahoana no nampiofanana an'i Shannon V1.5 hieritreritra tamin'ny fampiasana GRPO

1. Antony: Nahoana no Modely Fijerena?

Ho an'nyekipa mena AIfikarohana, ny fahatakaranany fombaahafahan'ny AI mamokatra vokatra mety hanimba dia zava-dehibe toy ny fandalinana ny vokatra ihany. Ny modely fijerena an'i Shannon V1.5 dia mampiseho ny fizotran'ny eritreritra manontolo, ahafahan'ny mpikaroka manao izao:

• Manara-maso ny lalan'ny fijerena mitarika ho amin'nyvokatry ny AI tsy voasivanafitondran-tena
• Mamantatra hoe aiza no tokony hidiran'ny fiarovana (mampisehony maha-zava-dehibe ny fiarovana AI)
• Mamolavola rafitra fiarovana matanjaka kokoa mikendry lamina fijerena manokana
• Mamorona tranga fitsapana mifanohitra tsara kokoa mifototra amin'ny fizotran'ny fijerena hita

2. Fahatakarana ny GRPO

Group Relative Policy Optimization (GRPO)dia fandrosoana mihoatra ny fomba RLHF mahazatra izay ahafahana mampiofana ny fahaiza-mieritreritra amin'ny fomba milamina sy mahomby kokoa. Novolavolain'ny DeepSeek AI, voaporofo fa tena mahomby amin'ny fampiofanana chain-of-thought izy.

Nahoana no GRPO fa tsy RLHF mahazatra?

Fototra Matematikan'ny GRPO

Ny GRPO dia manatsara ny politika amin'ny fampitahana ny valiny ao anatin'ny vondrona fa tsy amin'ny modely valisoa tanteraka:

Ity fampitahana mifandraika ity dia manana tombony maro:

• Fandaminana:Manitsy ho azy ny fahasamihafan'ny fahasarotana amin'ny bitsika samihafa
• Fahamarinana:Mampihena ny fahasamihafana amin'ny tombana gradient
• Fahombiazana:Tsy ilaina modely valisoa misaraka

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 Fandevonana

Mba hanombohana ny fahaiza-mieritreritra an'i Shannon V1.5, dia nampitantsika ny lamina chain-of-thought avy amin'ny modely fijerena an'i DeepSeek. Izany dia nanome fizotran'ny CoT avo lenta hampiofanana ny lohan'ny fijerena anay.

Famoronana Angon-drakitra DeepSeek

Fomba Fanangonana Soritra

Nanangona soritra fandinihana avy amin'ny sehatra maro samihafa izahay mba hiantohana ny fandrakofana feno ny fandinihana:

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. Maritrano Lohan'ny Fandinihana

Ny maodely Shannon V1.5 dia mampiditra iray manokanalohan'ny fandinihanaizay mamorona soritra fandinihana mazava alohan'ny vokatra farany. Ity fanampiny ara-maritrano ity dia ahafahana manao CoT mangarahara tsy manova ny maritrano Mixtral fototra.

Fampiharana ny Lohan'ny Fandinihana

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. Fizotry ny Fampiofanana

Dingana 1: Fampiofanana Mialoha ny Lohan'ny Fandinihana

Voalohany, manofana mialoha ny lohan'ny fandinihana amin'ny soritra CoT voasivana DeepSeek izahay amin'ny fampiasana fatiantoka cross-entropy mahazatra:

# 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

Dingana 2: Fanitsiana GRPO

Aorian'ny fampiofanana mialoha, mampihatra GRPO izahay mba hanatsarana ny kalitaon'ny fandinihana amin'ny fampiasana fampitahana mifandraika amin'ny vondrona:

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()
        }

Dingana 3: Fanatsarana Manokana ho an'ny Ekipa Mena

Farany, manitsy bebe kokoa amin'ny toe-javatra mpanohitra izahay mba hiantohana fa ny soritra fandinihana dia mampiseho tsara ny fandinihana ho an'nyAI tsy voasivana vokatry nyfanadihadiana:

6. Vokatra sy Fanadihadiana

Fepetra Kalitaon'ny Fandinihana

Fikarohana Lehibe

• Nihatsara be ny mangarahara:Avy amin'ny 12% ka hatramin'ny 97.8% amin'ny fandinihana no voalaza mazava ankehitriny
• Nitombo ny fahamarinan'ny fandinihana:Ny fandinihana mazava dia nanatsara ny kalitaon'ny valiny farany tamin'ny isa 12+
• Voamarina ny sandan'ny ekipa mena:Mitatitra ny mpikaroka momba ny fiarovana fa ny soritra fandinihana dia "sarobidy" amin'ny fahatakarana ny fandinihana fanararaotana
• Ny GRPO dia nihoatra ny RLHF:Isa fampifandraisana 15% tsara kokoa raha oharina amin'ny fomba nentim-paharazana