training_config.json

{
  "steps": 50000,
  "batch_size": 128,
  "grad_accum": 1,
  "seq_len": 4096,
  "d_model": 96,
  "n_layers": 12,
  "n_heads": 6,
  "n_kv_heads": 2,
  "chunk": 8,
  "cma_heads": 2,
  "expand": 2,
  "cma_identity_prob": 0.9,
  "lr": 0.0025,
  "muon_lr": 0.03,
  "muon_momentum": 0.95,
  "muon_ns_steps": 5,
  "muon_nesterov": true,
  "warmup": 750,
  "decay_start_frac": 0.5,
  "min_lr_ratio": 0.0,
  "wd": 0.01,
  "grad_clip": 1.0,
  "log_every": 5,
  "early_eval_step": 1000,
  "eval_every": 10000,
  "val_batch_size": 32,
  "val_batches": 0,
  "val_stride": 2048,
  "infer_tokens": 512,
  "infer_repeat_penalty": 1.2,
  "infer_prompt": "The process of photosynthesis",
  "lm_eval_tasks": "arc_easy,arc_challenge,hellaswag,piqa",
  "lm_eval_batch_size": "auto",
  "lm_eval_device": "cuda",
  "lm_eval_dtype": "float32",
  "lm_eval_softmax_dtype": "float32",
  "lm_eval_expected_version": "0.4.12",
  "lm_eval_retries": 3,
  "lm_eval_export_dir": "cma_1m_byte_lm_eval_hf",
  "lm_eval_output_dir": "lm_eval_results_cma_1m_byte",
  "arithmark_batch_size": 16,
  "arithmark_data_path": "benchmark_cache/arithmark_2.0.jsonl",
  "arithmark_force_download": false,
  "hf_repo_id": "User01110/cma-1M-exp",
  "hf_repo_private": false,
  "hub_upload_retries": 3,
  "data_seed": 1337,
  "shuffle_buffer": 50000,
  "tokenize_batch_size": 64,
  "compile": true
}