{
  "model_type": "sonamath",
  "architectures": [
    "SonaMathForCausalLM"
  ],
  "model_name": "SonaMath-0.5B",
  "brand": "SonaMath",
  "version": "0.1.1-sft-preview",
  "release_stage": "sft",
  "params_million": 591.9,
  "hidden_size": 1536,
  "num_attention_heads": 24,
  "num_key_value_heads": 4,
  "intermediate_size": 4352,
  "num_hidden_layers": 16,
  "max_position_embeddings": 2048,
  "vocab_size": 32000,
  "torch_dtype": "bfloat16",
  "language": "en",
  "license": "apache-2.0",
  "pipeline_tag": "text-generation",
  "tags": [
    "math",
    "reasoning",
    "research-preview",
    "sft",
    "small-language-models",
    "architecture-research",
    "sonamath"
  ],
  "training": {
    "pretrain_tokens_approx": "3B",
    "post_training": "light SFT only (this release)",
    "stage": "sft-preview",
    "rl_applied_to_this_checkpoint": false
  },
  "eval": {
    "gsm8k_full_greedy": {
      "accuracy": 0.1379833206974981,
      "correct": 182,
      "n": 1319,
      "parse_rate": 0.9931766489764974,
      "wilson_95_low": 0.12,
      "wilson_95_high": 0.158,
      "decoding": "greedy",
      "max_new_tokens": 1024,
      "applies_to": "SFT checkpoint only"
    }
  },
  "disclaimers": [
    "Weights are SFT-stage only.",
    "GSM8K score must not be cited as RL or anti-GRPO evidence.",
    "Former preview name: JunMath."
  ],
  "notes": "SonaMath SFT research preview. GSM8K full: 13.8% (182/1319), greedy, max_new=1024. Formerly branded JunMath.",
  "former_names": [
    "JunMath",
    "JunMath-0.5B"
  ]
}