Fix dead split parameter in PackedStreamingDataset._load_dataset

Refactor runner.py: extract shared setup logic into _setup_and_train helper

Fix dtype mismatch in RoPE cos/sin for mixed precision training

Replace F.silu with explicit SiLU implementation in SwiGLUFeedForward

Replace F.scaled_dot_product_attention with explicit implementation

Remove dead attn_dropout layer from GroupedQueryAttention

Fix LR warmup ordering and align adam_eps with Meta LLaMA

Remove redundant detect_scenario from LossDebugger

Add Code CPT pipeline for injecting Python code capability

Fix false/inaccurate citations in LossDebugger

Fix LR reference table and batch-LR scaling guidance in LossDebugger

Fix batch size diagnostic: widen window and list multiple causes

Fix gradient clipping thresholds in dynamics and checklist modules

Fix gradient diagnostic thresholds with evidence-based criteria in LossDebugger

Fix check_numerical_stability accuracy and completeness

Scale overfit test LR and steps by model size in LossDebugger

Reduce LOSS_BOUNCE false positives with moving-average smoothing

Improve LOSS_BOUNCE detection with pre-computed bounce metrics

Add LOSS_BOUNCE detection to diagnose_status classification chain

Add NaN detection to diagnose_status classification chain

Fix debugger reference data and scenario detection logic

Remove unused tokenizer training code (train_bpe, load_sentencepiece, load_trained_hf)

Add configurable wandb log directory path

Tighten expected loss ranges for FineWeb-Edu dataset

Use LLaMA 2 pretrained tokenizer and remove tokenizer_mode option

Fix BPE tokenizer ByteLevel decoder and update evaluation notebook

feat(training): add LossDebugger 5-level diagnostic framework

feat(config): add scale-specific presets to TrainConfig

fix(trainer): correct attribute name from total_mem to total_memory

refactor(runner): replace manual seed setup with set_seed utility

fix(device): correct attribute name from total_mem to total_memory

refactor(model): replace single-letter vars with descriptive names for readability

docs: translate all Korean comments and docstrings to English

refactor(data): replace per-worker seed strategy with full sharding in IterableDataset

Initial commit: LLM-1B-Lab project setup