a-426

This archive preserves the saved training_state.pt and all 24 rank RNG files from the final 4B Task149 checkpoint. The original save/load acceptance records are included. Inference weights and tokenizer are stored separately: https://huggingface.co/songdj/TraceGym-Qwen3.5-4B-GRPO-checkpoint-75

The matching model repository is private and requires its own access permission.

The 50,475,204,545-byte training state is split into six ordered pieces. After downloading this archive, run python restore_training_state.py --output /path/to/checkpoint-75/training_state.pt and copy the 24 rng-rank-*.pt files plus COMPLETE.json, MODEL_CHECKPOINT.json, and SAVE_ACCEPTANCE.json into that same checkpoint directory. Historical absolute paths in these original records describe the source machine and may need remapping. The script verifies the reconstructed state's length and SHA-256 before publishing its output filename. Use the model revision in state_manifest.json to obtain the matching model files. Reconstruction requires about 50.5 GB of additional disk space.

The state includes saved optimizer/master-weight/scheduler/runtime state. Resuming requires the compatible project training code, environment, curriculum and dataset; this archive does not itself launch or validate a new training run. Original 24-rank restore acceptance is historical evidence, not a new restore performed during upload.

Project: https://github.com/bbsngg/TraceGym/tree/codex/tracegym-grpo-completion-20260912