{
  "model_id": "MatterWaveAI/grt-30m-chat",
  "architecture": "GRTForCausalLM",
  "parameters": 29305498,
  "nn_linear_layers": 0,
  "chat_template": "ChatML (<|im_start|> / <|im_end|>)",
  "before_sft": {
    "val_loss": 5.2083,
    "val_ppl": 182.7854
  },
  "after_sft": {
    "val_loss": 0.3129,
    "val_ppl": 1.3674,
    "samples": [
      {
        "user": "Can you explain this briefly? What is GRT-30M?",
        "target": "GRT-30M is a 29.3M-parameter Givens Rotational Transformer developed by MatterWave AI that replaces every dense nn.Linear matrix with cascaded 2D Givens rotations and butterfly permutations.",
        "generated": "GRT-30M is a 29.3M-parameter Givens Rotational Transformer developed by MatterWave AI that replaces every dense nn.Linear matrix with cascaded 2D Givens rotations and butterfly permutations."
      },
      {
        "user": "Can you explain this briefly? How does a Givens Rotational Transformer replace dense linear layers?",
        "target": "Instead of O(D^2) dense matrix multiplications, GRT represents hidden states as planar coordinate pairs and transforms them through butterfly-permuted 2D Givens rotations modulated by Clifford algebra features.",
        "generated": "Instead of O(D^2) dense matrix multiplications, GRT represents hidden states as planar coordinate pairs and transforms them through butterfly-permuted 2D Givens rotations modulated by Clifford algebra features."
      },
      {
        "user": "Can you explain this briefly? What is PearlNorm in the GRT architecture?",
        "target": "PearlNorm is a pair-aware RMS normalization layer that divides each planar (x, y) pair by the root-mean-square of pair radii, rescaling magnitudes while preserving every pair's phase angle.",
        "generated": "PearlNorm is a pair-aware RMS normalization layer that divides each planar (x, y) pair by the root-mean-square of pair radii, rescaling magnitudes while preserving every pair's phase angle."
      },
      {
        "user": "Can you explain this briefly? What is RadialGELU?",
        "target": "RadialGELU applies a GELU activation to the Euclidean radius r of each 2D coordinate pair as gelu(r) / r, introducing nonlinearity on pair magnitudes without distorting phase directions.",
        "generated": "RadialGELU applies a GELU activation to the Euclidean radius r of each 2D coordinate pair as gelu(r) / r, introducing nonlinearity on pair magnitudes without distorting phase directions."
      },
      {
        "user": "Can you explain this briefly? How are token embeddings parameterized in GRT?",
        "target": "Each token in GRTEmbedding is stored as hidden_size // 2 phase angles plus a scalar log-magnitude, emitting exp(log_magnitude) * (cos(theta), sin(theta)) on the hypersphere.",
        "generated": "Each token in GRTEmbedding is stored as hidden_size // 2 phase angles plus a scalar log-magnitude, emitting exp(log_magnitude) * (cos(theta), sin(theta)) on the hypersphere."
      }
    ]
  }
}