---
license: apache-2.0
library_name: pytorch
datasets:
  - nur-dev/pns-world
tags:
  - persistent-memory
  - event-stream
  - reproducibility
---

# Persistent Event-State models

Assets for *Learning Persistent Referential Identity for Two-Hop Retrieval 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/` |
| Reviewer-requested tests | `rr2/campaign/`, 15 model-only `rr2/checkpoints/`, `rr2/reference/`, `rr2/frozen_evaluations/` |

## Reproduce reviewer-requested tests

Python 3.13 and a CUDA GPU for evaluation/training.

```bash
pip install huggingface_hub
hf download nur-dev/pns-bind-25m --revision paper-v2.1 --local-dir pns-model
hf download nur-dev/pns-world --repo-type dataset --revision paper-v2.1 --local-dir pns-data
cd pns-model
pip install -r e8_full/requirements.txt
python rr2/reproduce.py verify --data ../pns-data/rr2
CUDA_VISIBLE_DEVICES=0 python rr2/reproduce.py evaluate --data ../pns-data/rr2 --suite compose --arm future --seed 821 --split confirmation --output rr2_eval_821
```

Repeat with seeds `822`, `823`; use `--arm ordinary` for the ordinary-only control.
Use `--suite original --arm ordinary --split known_development` for the original-name
control, or `--suite gru --update 12000` for the semantic-only recurrent baseline.
Raw predictions are compared with published arrays. Checkpoints contain identical
model tensors with optimizer state omitted; hashes are in `rr2/INFERENCE_EXPORT.json`.

```bash
python rr2/reproduce.py prepare-training --data ../pns-data/rr2 --output rr2_train --workers 24
CUDA_VISIBLE_DEVICES=0 python rr2/reproduce.py train --data ../pns-data/rr2 --output rr2_train --suite compose --arm future --seed 821
```

Training corpora are regenerated deterministically and checked against frozen hashes.
`rr2/GPU_RESULTS.json` contains all primary and secondary criterion outcomes.

## Reproduce E8-FULL

```bash
python e8_full/reproduce.py verify --data ../pns-data/e8_full
CUDA_VISIBLE_DEVICES=0 python e8_full/reproduce.py evaluate --data ../pns-data/e8_full --arm grounded_future --seed 821 --output e8_eval_821
```

Repeat for arms `numeric_future`, `grounded_cut` and seeds `822`, `823`.

## Reproduce E1 / E3

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

`paper-v1.0` and `paper-v2.0` retain the original releases.