This directory is a complete evaluator and reproduction package for enrollment 25M2111.

The following is the shortest complete evaluator path. Run it from this Submission01 directory. The compressed artifact is restored first; do not pass the .ptz file directly to Transformers or vLLM.

Python 3.11 or later and a CUDA-capable machine are required. Any Conda or virtual environment may be used; its name is irrelevant. The following creates an isolated environment for a clean evaluation. If an equivalent environment is already active, skip the two environment-creation lines. If starting from a fresh machine, clone the private submission repository using an account that has been added as a collaborator:

The printed digest must equal the SHA256 in the table above.

Restoration is deterministic. restored_hf/submission_report.json records the restoration details.

The 0 is the physical GPU id. Multiple GPUs may be supplied as 0,1. The evaluation script clears any inherited CUDA mask, runs the checkpoint suite with max_new_tokens=65536 and repetition penalty 1.20, then merges shards. The keep_ids file is required because this submission evaluates logits over the 30,000 retained original output-token ids.

code.py provides both required functions:

The sparsity argument is accepted only for compatibility with the supplied course evaluator and is ignored by this compression method.

The command-line wrappers are convert_from_hf_checkpoint.py and dequantize_to_bf16.py. Evaluating the uploaded artifact requires only the dequantization wrapper. Rebuilding the artifact from the base model is the longer optional process below.

This reproduces the learned short-reasoning source, calibrated GPTQ weights, vocabulary selection, token-string predictor, and final lossless package. It requires the verified training/calibration trace JSONL and a GPU.

The produced sidecars are reproduced.ptz.json and reproduced.ptz.keep_ids.json.

The learned source is compressed with group-128 calibrated GPTQ: 3-bit MLP, 4-bit Gated DeltaNet projections, and 8-bit full attention. Sensitive recurrent controls and small tensors remain BF16. The tied embedding/output table stores 30,000 selected rows at 8 bits, and a 701,760-byte deterministic token-string predictor reconstructs omitted input rows. Zlib then losslessly reduces the serialized artifact from 1,790,943,184 to 1,626,481,790 bytes.

The physical restoration scored 0.879 accuracy, 0.970 parse rate, and 0.037 truncation rate on the 560-problem checkpoint suite at a recorded 32,768-token cap. Machine-readable configuration and evidence are in configs/ and results/ in this GitHub submission.