arch_check.json

{
  "learn_model": "luzimu/FullStack-Learn-LM-30B-A3B",
  "base_model": "Qwen/Qwen3-Coder-30B-A3B-Instruct",
  "architecture_comparison": {
    "architectures": {
      "learn": [
        "Qwen3MoeForCausalLM"
      ],
      "base": [
        "Qwen3MoeForCausalLM"
      ],
      "match": true
    },
    "hidden_size": {
      "learn": 2048,
      "base": 2048,
      "match": true
    },
    "num_hidden_layers": {
      "learn": 48,
      "base": 48,
      "match": true
    },
    "num_experts": {
      "learn": 128,
      "base": 128,
      "match": true
    },
    "num_experts_per_tok": {
      "learn": 8,
      "base": 8,
      "match": true
    },
    "num_attention_heads": {
      "learn": 32,
      "base": 32,
      "match": true
    },
    "num_key_value_heads": {
      "learn": 4,
      "base": 4,
      "match": true
    },
    "moe_intermediate_size": {
      "learn": 768,
      "base": 768,
      "match": true
    },
    "max_position_embeddings": {
      "learn": 262144,
      "base": 262144,
      "match": true
    },
    "model_type": {
      "learn": "qwen3_moe",
      "base": "qwen3_moe",
      "match": true
    }
  },
  "all_keys_match": true,
  "conclusion": "FullStack-Learn-LM-30B-A3B is a fine-tuned Qwen3-Coder-30B-A3B-Instruct (identical Qwen3Moe architecture)"
}