training_metadata.json

{
  "model": "logos23-gemma2_2b",
  "family": "gemma2_2b",
  "family_name": "Gemma 2 2B",
  "base_model": "google/gemma-2-2b",
  "base_model_quantized": "google/gemma-2-2b",
  "method": "LoRA (bf16)",
  "framework": "unsloth",
  "lora_rank": 64,
  "lora_alpha": 64,
  "lora_target_modules": [
    "q_proj",
    "k_proj",
    "v_proj",
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj"
  ],
  "epochs": 3,
  "effective_batch_size": 16,
  "load_in_4bit": false,
  "learning_rate": 0.0002,
  "lr_scheduler": "cosine",
  "max_seq_length": 2048,
  "dataset": "logos22_nothink.jsonl",
  "dataset_size": 895,
  "train_on_responses_only": true,
  "think_blocks": "stripped (no-think variant)",
  "final_loss": 1.2898975720717794,
  "runtime_seconds": 221.1209
}