# Replay-only decay tail continuation to recover the missing checkpoints after the original
# step_13500 decay-only run ran out of disk during checkpoint serialization.
# This resumes from the last valid saved checkpoint (`step_14200`) and reproduces the
# remaining original cooldown trajectory through `step_14850`.
#
# Important:
# - launch this replay with `--resume-mode model_only`
#   so the run keeps the model weights + global_step from `step_14200`
#   while intentionally resetting optimizer/scheduler to continue the missing tail
# - the original decay tail started at `step_13500`
# - original decay length was `1350` steps
# - therefore the correct tail end is `step_14850`
# - to continue the *same* inverse-proportional curve from `step_14200`, this replay
#   starts from the exact LR implied by the original tail at `step_14200`
#   and decays over the remaining `650` steps down to the same `final_lr`

resume_from: /mnt/dati/llm-nanochat/checkpoints_archive/tier2_20260629_medium_prune/20260628_resume-gpt2medium-gpt2preln-k20-wsddecayonly-lr2e-4-anchor20k-final2e5-webwiki-step13500/step_14200.pt
dataset_dir: /mnt/apps/llm-nanochat/datasets/202605141153_fineweb50_wiki50_50en_50it_score100_2500context_5Btokens_tok_20260515_en50it50_webwiki_stratified_500M
output_dir: /mnt/apps/llm-nanochat/artifacts/runs/20260629_resume-gpt2medium-gpt2preln-k20-wsddecayonly-rerunmissing-lr3p5294e5-anchor20k-final2e5-webwiki-step14200-to14850
tokenizer_dir: /mnt/apps/llm-nanochat/tokenizers/tokenizer_20260515_en50it50_webwiki_stratified_500M
seed: 1337

model:
  architecture: gpt2
  block_type: gpt2_prelayernorm
  tie_word_embeddings: true
  vocab_size: 32000
  dim: 1024
  n_layers: 24
  n_heads: 16

training:
  sequence_length: 2500
  max_steps: 14850
  batch_size: 2
  grad_accum_steps: 48

  learning_rate: 3.529411764705883e-05
  peak_lr: 3.529411764705883e-05
  lr_schedule: wsd-decay-only

  warmup_steps: 0
  stable_steps: 0
  decay_steps: 650
  final_lr: 2.0e-05

  adamw_betas:
    - 0.9
    - 0.95
  adamw_eps: 1.0e-08
  weight_decay: 0.1
  clip_grad_norm: 1.0

  save_every_steps: 50
  checkpoint_dir: /mnt/apps/llm-nanochat/checkpoints/20260629_resume-gpt2medium-gpt2preln-k20-wsddecayonly-rerunmissing-lr3p5294e5-anchor20k-final2e5-webwiki-step14200-to14850
  precision: bf16

evaluation:
  validation_every_steps: 50
  validation_max_batches: 128
  probe_every_steps: 1000
  probe_tokenizer_dir: /mnt/apps/llm-nanochat/tokenizers/tokenizer_20260515_en50it50_webwiki_stratified_500M
  probe_max_new_tokens: 32
  probe_prompts:
    en:
      - prompt: "The capital of Italy is"
        expected_next_text: " Rome"
      - prompt: "A small language model should"
        expected_next_text: " be"
    it:
      - prompt: "La capitale d'Italia è"
        expected_next_text: " Roma"
      - prompt: "Un piccolo modello linguistico dovrebbe"
        expected_next_text: " essere"