{
  "diagnostic_only": true,
  "model": "micro-llm",
  "benchmark": "WikiText-2",
  "instruction_dataset": "Stanford Alpaca held-out subset",
  "seed": 42,
  "protocol": {
    "pretraining_steps": 40,
    "post_training_steps": 24,
    "population": 4,
    "generations": 3,
    "dimensions": 128,
    "layers": 2,
    "sequence_length": 128,
    "bpe_vocab_size": 1024,
    "train_tokens": 4046002,
    "validation_tokens": 50000,
    "test_tokens": 50000,
    "parameters": 544256
  },
  "baseline": {
    "architecture": "gpt_baseline",
    "data_recipe": "wikitext",
    "post_training": "none",
    "validation_perplexity": 449.032,
    "test_perplexity": 416.134,
    "instruction_validation_loss": 6.3804
  },
  "champion": {
    "architecture": "parallel_gelu",
    "data_recipe": "wikitext",
    "post_training": "sft",
    "validation_perplexity": 433.336,
    "test_perplexity": 405.328,
    "instruction_validation_loss": 6.2309,
    "test_perplexity_reduction_percent": 2.60
  },
  "round_findings": {
    "structure": "parallel_gelu won the matched-budget structure round",
    "data": "WikiText-only beat 10/20/35 percent Tiny Shakespeare narrative mixtures on WikiText perplexity",
    "post_training": "plain SFT beat low-LR SFT and NEFTune alpha 5/10 under this small budget"
  },
  "checkpoint_committed": false,
  "raw_runs_committed": false
}