| { | |
| "model_type": "rodan-modern", | |
| "architecture": "ModernLM", | |
| "framework": "mlx", | |
| "stage": "chat", | |
| "base_model": "Rodan-10M (v9, PLE-free)", | |
| "params": 10410000, | |
| "vocab_size": 8194, | |
| "dim": 320, | |
| "n_layers": 8, | |
| "n_heads": 8, | |
| "n_kv_heads": 1, | |
| "head_dim": 40, | |
| "ffn_hidden": 768, | |
| "max_len": 512, | |
| "rope_base": 200000.0, | |
| "norm": "rmsnorm", | |
| "norm_eps": 1e-5, | |
| "activation": "swiglu", | |
| "qk_norm": true, | |
| "tied_embeddings": true, | |
| "value_residual": true, | |
| "ple_rank": 0, | |
| "lrm": true, | |
| "attention": "mqa", | |
| "chat_template": "chatml", | |
| "chat_tokens": {"im_start": 8192, "im_end": 8193}, | |
| "eot_id": 0, | |
| "tokenizer": "byte-level BPE (8k) + 2 ChatML specials = 8194", | |
| "recommended_decode": "greedy + repetition_penalty 1.3 + no-repeat-3gram (tiny models loop under pure greedy)", | |
| "notes": "Warm-started from Rodan-10M v9 (PLE-free). Instruction fold: smol-smoltalk ChatML + 45% curated replay (continued-pretrain, not masked SFT). Load with model_opt.ModernLM(ModernConfig(**fields)) + load_weights('model.safetensors'). Wrap prompts in ChatML: <|im_start|>user\\n{q}<|im_end|>\\n<|im_start|>assistant\\n" | |
| } | |