{
  "model_name": "Grug-12B",
  "base_model": "google/gemma-4-12B-it",
  "artifact_type": "peft_lora_adapter",
  "teacher_model_for_compact_reasoning": "cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit",
  "teacher_runtime": "vllm",
  "dataset_counts": {
    "normalized_recent_reasoning": 8680,
    "selected_verbose": 6144,
    "compact_raw": 6144,
    "compact_verified": 5740,
    "compact_rejected": 404,
    "packed_train": 5166,
    "packed_validation": 287,
    "packed_test": 287
  },
  "verified_source_counts": {
    "hotdogs/uka-glm-5.2": 1617,
    "Scale-or-Reason/general-reasoning-ift-pairs": 1305,
    "samcheng0/lumia-reasoning-sft-v1": 1103,
    "HSH-Intelligence/verified-math-reasoning-3k": 672,
    "kd13/CodeDebug-Instruct-v2-Reasoning": 600,
    "Madarabr/cortex-adaptive-thinking": 300,
    "CL-From-Nothing/code_rose_initial_1_7B_SFT_10K_rollouts_Qwen3-4B-Thinking-2507_k12_t0.7_maxtok12288": 143
  },
  "training": {
    "method": "QLoRA completion-only SFT",
    "quantization": "4-bit NF4, BF16 compute",
    "max_seq_length": 6144,
    "max_steps": 100,
    "batch_size": 1,
    "gradient_accumulation_steps": 16,
    "learning_rate": 0.00008,
    "lora_r": 16,
    "lora_alpha": 32,
    "lora_dropout": 0.05,
    "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    "final_eval_loss": 0.8895471096038818,
    "train_runtime_seconds": 2121
  },
  "local_eval": {
    "name": "EOS-only math proxy eval",
    "generation_mode": "eos_only",
    "max_new_tokens": null,
    "timeout_seconds": 900,
    "rows": 36,
    "base": {
      "total_generated_tokens": 8227,
      "avg_generated_tokens": 228.52777777777777,
      "proxy_accuracy": 0.9166666666666666,
      "contains_expected_rate": 0.9166666666666666,
      "numeric_last_match_rate": 0.8611111111111112
    },
    "adapter": {
      "total_generated_tokens": 2482,
      "avg_generated_tokens": 68.94444444444444,
      "proxy_accuracy": 1.0,
      "contains_expected_rate": 1.0,
      "numeric_last_match_rate": 1.0
    }
  },
  "acknowledgement": "Thanks to Lambda (https://lambda.com/), the inference provider, for compute credits."
}