training_info.json

{
  "model": {
    "base_model": "zai-org/GLM-4.5-Air",
    "final_model_path": "outputs_fsdp/final_model"
  },
  "training_config": {
    "lora_r": 128,
    "lora_alpha": 256,
    "lora_dropout": 0.05,
    "lora_target_modules": [
      "q_proj",
      "k_proj",
      "v_proj",
      "o_proj"
    ],
    "learning_rate": 2.5e-05,
    "lr_scheduler_type": "cosine",
    "micro_batch_size": 1,
    "gradient_accumulation_steps": 2,
    "effective_batch_size": 32,
    "sequence_length": 16384,
    "chunk_overlap": 2048,
    "weight_decay": 0.01,
    "max_grad_norm": 1.0,
    "warmup_ratio": 0.1,
    "eval_split": 0.05,
    "bf16": true,
    "seed": 42
  },
  "hardware": {
    "num_gpus": 16,
    "gpu_name": "NVIDIA H200",
    "num_nodes": 1,
    "gpus_per_node": 8
  },
  "phases": [
    {
      "phase": 1,
      "name": "phase1_foundation",
      "description": "Foundation: Learn codebase structure and file patterns",
      "dataset": "dataset/phase1_foundation.jsonl",
      "epochs": 2,
      "learning_rate": 2.5e-05,
      "warmup_ratio": 0.15,
      "num_train_samples": 9293,
      "num_eval_samples": 512,
      "num_chunks": 9805,
      "train_metrics": {
        "train_runtime": 45748.92132782936,
        "train_runtime_minutes": 762.4820221304893,
        "train_steps": 581,
        "train_loss": 0.5921854273129171,
        "train_perplexity": 1.8079352121008547,
        "samples_per_second": 0.40626094475136876,
        "steps_per_second": 0.012699752980767526
      },
      "eval_metrics": {
        "eval_loss": 0.36529209305808763,
        "eval_perplexity": 1.4409348337482015,
        "eval_accuracy": 88.77101374493351,
        "best_eval_loss": 0.36561795309899026
      }
    },
    {
      "phase": 2,
      "name": "phase2_evolution",
      "description": "Evolution: Learn commit patterns and code changes",
      "dataset": "dataset/phase2_evolution.jsonl",
      "epochs": 2,
      "learning_rate": 2e-05,
      "warmup_ratio": 0.1,
      "num_train_samples": 16622,
      "num_eval_samples": 1545,
      "num_chunks": 18167,
      "train_metrics": {
        "train_runtime": 88820.11419820786,
        "train_runtime_minutes": 1480.3352366367976,
        "train_steps": 1039,
        "train_loss": 0.790716471444525,
        "train_perplexity": 2.204975662547615,
        "samples_per_second": 0.37428458970243894,
        "steps_per_second": 0.0116978007670808
      },
      "eval_metrics": {
        "eval_loss": 2.551615942151948,
        "eval_perplexity": 12.827816051917177,
        "eval_accuracy": 40.84345327062199,
        "best_eval_loss": 2.5516352893463
      }
    },
    {
      "phase": 3,
      "name": "phase3_pr_mastery",
      "description": "PR Mastery: Learn PR review patterns and discussions",
      "dataset": "dataset/phase3_pr_mastery.jsonl",
      "epochs": 1,
      "learning_rate": 1.5e-05,
      "warmup_ratio": 0.05,
      "num_train_samples": 9797,
      "num_eval_samples": 509,
      "num_chunks": 10306,
      "train_metrics": {
        "train_runtime": 24744.46716451645,
        "train_runtime_minutes": 412.40778607527415,
        "train_steps": 306,
        "train_loss": 0.49508867293498876,
        "train_perplexity": 1.6406437133004639,
        "samples_per_second": 0.3959268928631,
        "steps_per_second": 0.012366400859049565
      },
      "eval_metrics": {
        "eval_loss": 0.5012174650255474,
        "eval_perplexity": 1.6507297535648182,
        "eval_accuracy": 90.171501333015,
        "best_eval_loss": 0.5012283607793506
      }
    }
  ],
  "phase_checkpoints": [
    "outputs_fsdp/phase1_foundation/final",
    "outputs_fsdp/phase2_evolution/final",
    "outputs_fsdp/phase3_pr_mastery/final"
  ],
  "summary": {
    "initial_train_loss": 0.5921854273129171,
    "final_train_loss": 0.49508867293498876,
    "initial_eval_loss": 0.36529209305808763,
    "final_eval_loss": 0.5012174650255474,
    "initial_perplexity": 1.4409348337482015,
    "final_perplexity": 1.6507297535648182,
    "total_epochs": 5,
    "total_phases": 3,
    "total_steps": 1926,
    "total_training_time_seconds": 161561.2551908493,
    "total_training_time_hours": 44.87812644190259
  },
  "timestamp": "20251211_212051",
  "run_name": "glm-air-curriculum-16gpu",
  "output_directory": "outputs_fsdp"
}