Security and Limitations

Security, Reproducibility and Limitations

Public-repository hygiene

• No API keys, credentials or access tokens should be committed. This repository does not require an OpenAI API key.
• Review local corpora for confidential, licensed or personal information before publishing.
• Model checkpoints and adapter weights may be large and contain learned information from training samples; publish them deliberately, not by default.
• Pretrained Qwen weights are downloaded from Hugging Face on demand. Review the relevant model card and license before redistribution.
• This source ZIP does not include a software license. Add one only after making an intentional licensing decision.

Technical limitations

• Full model replica per worker; no sharded optimizer/model state (no FSDP/ZeRO).
• Built-in corpus is small and educational. No validation split or meaningful generation benchmark.
• In Qwen mode loss covers all non-padding tokens; response-only masking is absent.
• Checkpointing happens only at the end, rank 0; optimizer/rank RNG state is not saved for resume.
• Failure of one worker can stop the distributed collective; no elastic recovery policy is implemented.
• Training data resides in process memory; no network data ingestion or distributed streaming storage.
• CUDA/NCCL and Qwen adapter mode were not validated in this package.

See Future-Enhancements.