# Two additional data passes from the frozen V2.0 best weights. # Interrupted step-zero artifacts are archived; new runs use a fresh output path. architecture = "tiny_gpt_v2" token_dir = "artifacts/token-data/corpus-v2-16k" index_dir = "artifacts/training-data/corpus-v2-c128" output_dir = "artifacts/models/tiny-ja-v2.1-e16k-d320-l6-continue" log_dir = "runs/training/tiny-ja-v2.1-e16k-d320-l6-continue" source_provenance = "handoff/source-provenance-v21.json" [initialization] checkpoint = "artifacts/models/tiny-ja-v2.0-e16k-d320-l6/best.pt" [runtime] compile_backbone = true [model] vocab_size = 16384 context_length = 128 d_model = 320 n_heads = 5 n_layers = 6 d_ff = 832 dropout = 0.0 norm = "rmsnorm" norm_eps = 0.00001 activation = "swiglu" bias = false tie_embeddings = true position_embedding = "learned" [training] device = "cuda" precision = "bf16" seed = 42 cpu_threads = 4 batch_size = 256 num_workers = 4 bucket_multiplier = 64 gradient_accumulation = 1 epochs = 2 data_epoch_offset = 4 max_steps = 0 learning_rate = 0.0003 min_learning_rate = 0.00003 warmup_steps = 500 weight_decay = 0.1 beta1 = 0.9 beta2 = 0.95 grad_clip = 1.0 prefix_crop_probability = 0.30 prefix_crop_min_remaining_tokens = 8 log_every = 100 eval_every = 5000 eval_batches = 64 checkpoint_every = 5000 benchmark_warmup_steps = 100 validate_full_at_end = true validate_full_each_epoch = true early_stopping_patience = 2 early_stopping_min_delta = 0.01 [evaluation] tokenizer_dir = "artifacts/tokenizers/ja-unigram-16k-v2" [evaluation.benchmarks] ajimee = "artifacts/benchmarks/ajimee-jwtd-v2-v1" development = "artifacts/benchmarks/ime-dev-v2" [verification] mode = "metadata" [tracking] enabled = true project = "vimeml" mode = "online"