| { | |
| "architectures": ["SealGlazerLM"], | |
| "model_type": "sealglazer", | |
| "vocab_size": 8192, | |
| "hidden_size": 128, | |
| "num_hidden_layers": 4, | |
| "num_attention_heads": 4, | |
| "intermediate_size": 384, | |
| "max_position_embeddings": 256, | |
| "rms_norm_eps": 1e-6, | |
| "tie_word_embeddings": true, | |
| "dtype": "float32", | |
| "rope_theta": 10000.0, | |
| "num_parameters": 1901696, | |
| "notes": "From-scratch Llama-style causal LM (RMSNorm + RoPE + SwiGLU MLP + MHA). Custom architecture, not a standard transformers model_type; load with the provided load_sealglazer.py. Embeddings are tied (single tok.weight used for both input embedding and lm_head)." | |
| } |