| # Historical Training Scripts | |
| These are the original launch scripts. Machine-specific absolute paths are deliberately | |
| preserved as provenance and must be reviewed before reuse. | |
| - `pretrain/codegpt_multilingual_5epoch.sh`: two-GPU llm.c CodeGPT-Multilingual CPT. | |
| - `pretrain/gpt2_124m_code_5epoch.sh`: two-GPU llm.c GPT-2 124M CPT. | |
| - `pretrain/qwen25_coder_7b_full_4gpu.sh`: four-GPU, full-parameter Qwen2.5-Coder-7B CPT. | |
| - `pretrain/launch_and_log.sh`: captures the ms-swift command, environment, GPU record, | |
| package freeze, and log used by the 7B run. | |
| - `pretrain/stage1_short.sh`: Qwen2.5-Coder-0.5B Stage-1 CPT. | |
| - `sft/run_qwen25_coder_0p5b_instruct_F3_v2_1gpu.sh`: full-parameter 0.5B F3 SFT. | |
| For llm.c, use `code/llmc_training/` and the exact Git commit recorded in the run evidence. | |
| For ms-swift, compare each checkpoint's `args.json` with its launch script before rerunning. | |