{
  "model_type": "log_linear_gdn2_370M",
  "arch": "log-linear GDN-2 (fla chunk_log_linear_attn)",
  "tokens_trained": 999292928,
  "global_batch_tokens": 2097152,
  "micro_batch_size": 32,
  "grad_accum": 2,
  "seq_len": 4096,
  "num_params": "370M",
  "tokenizer": "TinyLlama/TinyLlama_v1.1",
  "corpus": "fineweb-edu",
  "note": "global_batch=2M (4x larger than vanilla/SSC/GRM 0.5M). Faster eval only, not directly comparable to vanilla 1B."
}