run_metrics.json

{
  "run_name": "llama32-3B-rte-bf16-lora-seed42",
  "model_size": "3B",
  "model_id": "meta-llama/Llama-3.2-3B",
  "task": "rte",
  "glue_config": "rte",
  "n_classes": 2,
  "bitwidth": "bf16",
  "seed": 42,
  "lora_init_source": "shared:lora_init_3B_seed42.pt:224tensors",
  "trainable_params": 9175040,
  "total_params": 3221924864,
  "train": {
    "steps": 105,
    "epochs": 2.9411764705882355,
    "train_runtime_sec": 225.48081922531128,
    "train_loss": 0.4097530353636969,
    "n_train_examples": 2241,
    "n_train_tokens_per_epoch": 219447,
    "throughput_samples_per_sec": 29.231650360477072,
    "throughput_tokens_per_sec": 2862.4712077892063
  },
  "peak_gpu_mem_gib": {
    "allocated": 7.778398513793945,
    "reserved": 12.90234375
  },
  "validation": {
    "accuracy": 0.8755020080321285,
    "macro_f1": 0.8754698717431637,
    "loss": 0.32502645490519017,
    "n": 249,
    "n_classes": 2,
    "majority_baseline": 0.5140562248995983,
    "pred_dist": {
      "A": 124,
      "B": 125
    },
    "true_dist": {
      "A": 121,
      "B": 128
    }
  },
  "test": {
    "accuracy": 0.8303249097472925,
    "macro_f1": 0.8300058758242475,
    "loss": 0.388077862426262,
    "n": 277,
    "n_classes": 2,
    "majority_baseline": 0.5270758122743683,
    "pred_dist": {
      "A": 143,
      "B": 134
    },
    "true_dist": {
      "A": 146,
      "B": 131
    }
  },
  "hyperparams": {
    "num_train_epochs": 3,
    "learning_rate": 0.0002,
    "max_length": 256,
    "per_device_train_batch_size": 4,
    "per_device_eval_batch_size": 8,
    "gradient_accumulation_steps": 16,
    "warmup_ratio": 0.03,
    "lr_scheduler_type": "cosine",
    "gradient_checkpointing": true
  },
  "lora": {
    "r": 16,
    "alpha": 32,
    "dropout": 0.0,
    "bias": "none",
    "target_modules": [
      "q_proj",
      "k_proj",
      "v_proj",
      "o_proj"
    ]
  },
  "env": {
    "torch": "2.4.1+cu121",
    "cuda": "12.1",
    "gpu": "NVIDIA GeForce RTX 4090",
    "slurm_job": "1932445",
    "node": "node39"
  },
  "debug_caps": {
    "max_train": null,
    "max_eval": null,
    "max_steps": -1
  }
}