logging stdout/stderr to research/results/embedding_init/gemma_mapping_byteify_cpt_seq1024/train.log seed=42 flash SDP TF32 using measured preliminary seq=1024 layout batch=16 accum=8; CPT lock remains seq=2048 train layout batch=16 accum=8 (128 seq/step, 131072 tok/step) checkpointing=off liger=True loading tokenizer research/results/embedding_init/gemma_mapping_byteify full token census over local corpus using full token census from /workspace/ugrip-kd/research/scripts/embedding_initialization/results/corpus_token_counts.json (cached tokenizer /workspace/ugrip-kd/research/results/expansion/tokenizer; current research/results/embedding_init/gemma_mapping_byteify) am census: 476898623 tokens (1959.7 tok/doc, 243349 docs) ta census: 4369273220 tokens (924.0 tok/doc, 4728460 docs) kk census: 4464108618 tokens (1632.8 tok/doc, 2733982 docs) tl census: 306736374 tokens (880.3 tok/doc, 348453 docs) en census: 1542145528 tokens (771.1 tok/doc, 2000000 docs) splits am: full 243349 eval 256 pool 243093 ta: full 4728460 eval 256 pool 4728204 kk: full 2733982 eval 256 pool 2733726 tl: full 348453 eval 256 pool 348197 en: full 2000000 eval 256 pool 1999744 mix weights {'am': 0.0446, 'ta': 0.4089, 'kk': 0.4178, 'tl': 0.0287, 'en': 0.1} (corpus-proportional; no document repeat) lang token budgets {'am': 49803264, 'ta': 456292352, 'kk': 466195456, 'tl': 32032768, 'en': 111591424} packing via HuggingFace datasets.map (batched, num_proc; not cached) am: datasets.map n=35332 batched=1024 num_proc=8 hf-map am (num_proc=8): 0%| | 0/35332 [00:00