DASE7506 MP2 frozen weights
This repository provides frozen inference weights and configurations from three rounds of DASE7506 MP2 experiments. Full code, course data, training and evaluation records are in KokoNAa/7506_m2. The matching code snapshot is b091741.
Each model selects one of four candidate solutions for a math question. Inference uses a custom NumPy interface with course-provided fixed embeddings and candidate-text features. These weights do not use the Transformers from_pretrained interface.
Files and results
| Version | Weight directory | Validation | Full public test |
|---|---|---|---|
| Round 1: pairwise ranking | release/selected_model/ |
741/798, 92.86% | 380/467, 81.37% |
| Round 2: linear/MLP blend | release/round2_candidate/ |
745/798, 93.36% | 386/467, 82.66% |
| Round 3: additional numeric/arithmetic features | release/round3_candidate/ |
746/798, 93.48% | 389/467, 83.30% |
Each directory contains matching model.npz and config.json files. Round-three inference assets total 287282 bytes. SHA256SUMS lists SHA-256 checksums for all six assets. bundle_manifest.json records the code version, frozen-manifest hashes, file sizes and official scores.
Relative to round two, round three gains 3 correct test answers, or 0.64 percentage points: 16 questions are fixed and 13 regress. This is a small observed improvement, rather than evidence of a stable or statistically significant gain.
Download and reproduce round three
Install Git LFS and Git Xet, then run the following from a common parent directory:
git xet install
git clone https://github.com/KokoNAa/7506_m2.git
git clone https://huggingface.co/Fengqinfeng/7506_mp2
cp -R 7506_mp2/release/. 7506_m2/release/
cd 7506_m2
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install -r code/requirements.txt
cd code
export OMP_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2 MKL_NUM_THREADS=2
export NUMEXPR_NUM_THREADS=2 VECLIB_MAXIMUM_THREADS=2 BLIS_NUM_THREADS=2
python predict.py --method submission_round3.py \
--checkpoint ../release/round3_candidate \
--split test --output runs/reproduced-round3-predictions.json
python score_predictions.py --split test \
--predictions runs/reproduced-round3-predictions.json \
--output runs/reproduced-round3-score.json
From this weights repository root, run shasum -a 256 -c SHA256SUMS to verify downloaded files. For round two, use submission_round2.py and ../release/round2_candidate; for round one, use submission.py and ../release/selected_model.
Evaluation protocol and provenance
The protocol is dase7506-mp2-open-hard-verifier-v5. Training uses only the supplied 4755 training questions; validation is used for development and frozen selection. The complete public test set contains 467 questions. The official scorer checks the selected candidate's final numeric answer, rather than every reasoning step. No external datasets or additional pretrained model weights are used.
All published assets match the pre-test freeze manifests byte for byte. Publication does not retrain or tune any model. Full prediction plus independent scoring takes 2.24 seconds for round three and 1.94 seconds for round two on the same server. Both use two threads; the server's actual CPU-time quota during measurement is approximately 0.5 cores. These timings describe that server; the course reference machine determines final resource eligibility.
Original course project: xudongwu-0/DASE7506-MP2. Full attribution and data licensing are retained in the code repository. Codex assisted with experiment design, implementation, execution, analysis and publication records.