{
  "track": "A - Extended Pre-Training",
  "repo_name": "verl",
  "model": "/root/.cache/huggingface/hub/models--Qwen--Qwen2.5-Coder-3B/snapshots/09d9bc5d376b0cfa0100a0694ea7de7232525803",
  "corpus_dir": "data/code_corpus_verl",
  "extensions": [
    ".py"
  ],
  "split_mode": "file",
  "eval_split": "test",
  "eval_block_size": 1024,
  "corpus_files": 214,
  "train_files": 172,
  "val_files": 21,
  "test_files": 21,
  "val_file_names": [
    "verl__workers__actor__dp_actor.py",
    "verl__utils__reward_score__prime_math____init__.py",
    "verl__models__mcore__registry.py",
    "verl__trainer__ppo__utils.py",
    "verl__workers__config__model.py",
    "verl__checkpoint_engine__hccl_checkpoint_engine.py",
    "verl__utils__memory_buffer.py",
    "verl__utils__config.py",
    "verl__utils__py_functional.py",
    "verl__workers__rollout__utils.py",
    "verl__models__llama__megatron__layers__parallel_attention.py",
    "verl__trainer__ppo__rollout_corr_helper.py",
    "verl__trainer__config__algorithm.py",
    "verl__utils__net_utils.py",
    "verl__utils__profiler__config.py",
    "verl__utils__fs.py",
    "verl__trainer__sft_trainer_ray.py",
    "verl__workers__rollout__vllm_rollout__utils.py",
    "verl__models__weight_loader_registry.py",
    "verl__utils__tracking.py",
    "verl__trainer__ppo__core_algos.py"
  ],
  "test_file_names": [
    "verl__workers__rollout__tokenizer.py",
    "verl__models__mcore__qwen2_5_vl__attention.py",
    "verl__protocol.py",
    "verl__utils__transformers_compat.py",
    "verl__workers__rollout__sglang_rollout__http_server_engine.py",
    "verl__utils__import_utils.py",
    "verl__models__qwen2__megatron__layers__parallel_linear.py",
    "verl__models__qwen2__megatron__layers__parallel_rmsnorm.py",
    "verl__models__qwen2__megatron__checkpoint_utils__qwen2_saver.py",
    "verl__workers__critic__dp_critic.py",
    "verl__checkpoint_engine__nccl_checkpoint_engine.py",
    "verl__utils__megatron__optimizer.py",
    "verl__models__qwen2__megatron__layers__parallel_decoder.py",
    "verl__models__llama__megatron__checkpoint_utils__llama_loader_depracated.py",
    "verl__utils__logging_utils.py",
    "verl__trainer__ppo__reward.py",
    "verl__interactions__base.py",
    "verl__models__mcore__bridge.py",
    "verl__workers__config__actor.py",
    "verl__single_controller__base__worker_group.py",
    "verl__utils__memory_utils.py"
  ],
  "block_size": 1024,
  "sequence_curriculum": [
    768,
    1024
  ],
  "train_chunks": 533,
  "val_chunks": 84,
  "test_chunks": 51,
  "epochs": 3.0,
  "batch_size": 8,
  "eval_batch_size": 8,
  "grad_accum_steps": 1,
  "learning_rate": 0.0001,
  "lora_r": 16,
  "baseline_validation_perplexity": 3.182,
  "baseline_test_perplexity": 2.7764,
  "baseline_perplexity": 2.7764,
  "post_training_validation_perplexity": 2.7844,
  "post_training_test_perplexity": 2.2379,
  "post_training_perplexity": 2.2379,
  "perplexity_reduction": 0.5385,
  "improvement_percent": 19.3955,
  "training_loss": 0.9615,
  "stage_summaries": [
    {
      "stage": 1,
      "block_size": 768,
      "epochs": 1.5,
      "train_chunks": 711,
      "val_chunks": 112,
      "test_chunks": 0,
      "training_loss": 1.0578,
      "validation_loss": 1.1055
    },
    {
      "stage": 2,
      "block_size": 1024,
      "epochs": 1.5,
      "train_chunks": 533,
      "val_chunks": 84,
      "test_chunks": 0,
      "training_loss": 0.8652,
      "validation_loss": 1.023
    }
  ]
}