25M2111-Week02-Track2-40-Submission01

Week 02 — Track 2 — 40% target

This is the complete evaluation and reproduction package for enrollment 25M2111. The selected checkpoint is the Round-18 tail-Fisher block-adaptive model, chosen as a transparent post-holdout robustness override.

Evaluate the uploaded checkpoint

Run from this Submission01 directory:

python -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[eval]'

hf download safffrron/25M2111-Week02-Track2-40-Submission01 \
  --local-dir hf_checkpoint
sha256sum hf_checkpoint/week02_40_tail_fisher_block64.pt

python dequantize_to_bf16.py \
  Qwen/Qwen3.5-4B hf_checkpoint restored_hf
python inference/smoke.py restored_hf
./evaluation/run_public_eval.sh restored_hf 0

The digest must match the table. Evaluation defaults to max_new_tokens=65536 and repetition penalty 1.20; recorded artifact evidence used 32,768 tokens.

Required conversion API

code.py provides:

• convert_from_hf_checkpoint(model_name, output_path, sparsity=None)
• convert_to_hf_checkpoint(model_name, checkpoint_path, output_path)

The optional sparsity argument is ignored because this is mixed-precision quantization, not pruning. checkpoint_path may be the download directory or the artifact file. Missing files are fetched from the submission repository.

Method, selection, and reproduction

The short-reasoning source is partitioned into 64-row blocks. A nested successive-refinement allocator preserves the Round-16 widths and spends the remaining budget by upgrading 212 answer-tail-Fisher-ranked blocks (36,012,032 weights) from 6 to 8 bits. The physical stream contains dense mixed-width integer codes, FP16 group scales, byte selectors, and protected BF16 tensors.

The frozen checkpoint-only protocol selected the Week-1 block64 model at 0.929 versus this model's 0.918. We nevertheless chose this artifact after it won the sealed 2026 holdout 0.603 versus 0.492, with 8/1 paired flips (p=0.039062). This is documented as a post-holdout robustness decision, not as the precommitted winner.

For a full rebuild, train the short source with training/reproduce_source.sh, set EAIMATH_BLOCK64_SOURCE to it and EAIMATH_BLOCK64_REPORT to the downloaded block_adaptive_report.json, then run convert_from_hf_checkpoint.py. Exact results are in results/.

Track 2 CUDA rationale

This is byte-for-byte the corresponding Track 1 representation. CUDA_ACCELERATION.md explains how its physical operators map to packed CUDA execution and clearly separates that realizability argument from the submitted BF16-compatible evaluation path.