Environment Record and Known Reproducibility Limitations

Run-Specific Evidence Available

The Qwen2.5-Coder-7B run has a contemporaneous package freeze, command, environment and GPU snapshot under evidence/training_runs/qwen25_coder_7b_cpt/. Key versions were:

Component	Version
ms-swift	4.0.2
PyTorch	2.5.1+cu121
Transformers	5.3.0
Accelerate	1.13.0
DeepSpeed	0.18.8
FlashAttention	2.8.3, CUDA 12 / PyTorch 2.5 / CPython 3.11 wheel
GPUs	four NVIDIA H100 80 GB devices

The run completed 18,165/18,165 steps and three epochs with recorded train loss 0.5602 and token accuracy 0.8744. These are training telemetry, not downstream benchmark scores.

The GPT-2 and CodeGPT runs record the exact Docker image nvidia/cuda:12.2.2-devel-ubuntu22.04, build command, effective trainer arguments, and the clean llm.c Git commit. They do not include a Python package freeze because the trainer was built and executed inside that CUDA development image.

Known Gaps

The following May 2026 runs do not have their own complete package or system snapshots:

models/pretrained/stage1_cpt
models/sft/sft_f3_refined_instruct

Their checkpoint args.json files record effective ms-swift parameters, and their exact datasets and historical launch scripts are included. A nearby April 2026 ms-swift snapshot used the same key versions as the 7B run, but treating that as the exact May environment is an inference. Bit-for-bit numerical retraining is not claimed.

The shared 20 GB raw dataset used by CodeGPT, GPT-2 and Qwen2.5-Coder-7B is not duplicated in this package. Its historical source path, size and record count are recorded in the main README. Reproducing those runs requires authorised access to that exact source data.

Raw llm.c model checkpoints and optimiser states are excluded after conversion to standard Hugging Face checkpoints. Qwen optimiser, scheduler and RNG states are also excluded. The delivered weights support inference and further training, but not exact mid-run resumption.

Delivery Validation Environment

Validation uses the locally installed Python, PyTorch, Transformers and Safetensors versions recorded in evidence/smoke_test_results.json. The smoke test checks that each checkpoint is structurally loadable and produces finite, non-degenerate logits; it is not a scientific performance evaluation.