nur-dev/pns-bind-25m

Persistent Event-State models

Assets for Learning Persistent Referential Identity for Multi-Step Memory in Event-Stream Models.

Experiment	Assets
E1 / E3	src/, eval/, configs/, checkpoints/, results/, reproduce/
E8-FULL	e8_full/source/, nine final e8_full/checkpoints/, e8_full/results/, e8_full/reference/

Reproduce E8-FULL

Python 3.13; CUDA GPU for evaluation. Download both repositories at paper-v2.0:

pip install huggingface_hub
hf download nur-dev/pns-bind-25m --revision paper-v2.0 --local-dir pns-model
hf download nur-dev/pns-world --repo-type dataset --revision paper-v2.0 --local-dir pns-data
cd pns-model
pip install -r e8_full/requirements.txt
python e8_full/reproduce.py verify --data ../pns-data/e8_full
python e8_full/reproduce.py evaluate --data ../pns-data/e8_full --arm grounded_future --seed 821 --output e8_results/grounded_future_821

Repeat evaluation for arms grounded_future, numeric_future, grounded_cut and seeds 821, 822, 823. The default split is confirmation; use --split development for development. --smoke checks one batch per panel. Results are compared row by row with the published outcomes. summarize reproduces reported metrics and gates without a GPU.

Reproduce E1 / E3

pip install -r requirements.txt
PNS_DATA=../pns-data bash reproduce/reproduce_headline.sh
bash reproduce/reproduce_tables.sh

paper-v1.0 retains the original E1/E3 release. Data and split manifests are in nur-dev/pns-world.