This repository is an automated off-node backup of Qwen/Qwen2.5-7B-Instruct-derived MemAgent-PTE r131 local step 20; initialized from r107 step 210 with optimizer restart. The latest verified export corresponds to global step 20.

The checkpoint includes custom modeling code. Load it only after reviewing that code, and pass trust_remote_code=True to Transformers. The exported safetensor payload is stored in float32; select an appropriate lower-precision torch_dtype at load time when GPU memory is constrained.

This is a research checkpoint derived from the Apache-2.0 Qwen model named in the source label. It was trained for long-context memory behavior and has not received a general safety evaluation. Validate it for the intended task before deployment. The training and uploader code is maintained at https://github.com/ZhangAIPI/mem-agent-pte.

backup_manifest.json records the exact checkpoint step, source metadata, file sizes, and upload time. During training, main is replaced only after a durable local checkpoint has been validated. Intermediate history is squashed after verification so this repository remains a latest-checkpoint backup.