# GraphBSI checkpoints

| Dataset | File | Source run |
|---------|------|------------|
| Moses | moses/last.ckpt | u2xo7qql |
| GuacaMol | guacamol/last.ckpt | 635bennr (epoch 89) |
| GuacaMol (50-step Euler row of Tab. 1) | guacamol/epoch_59.ckpt | 17oqxx3h (epoch 59) |
| Planar | planar/last.ckpt | agag753c |
| SBM | sbm/last.ckpt | aog3lu3i |
| Tree | tree/last.ckpt | n6t28f9t |
| QM9 (no H) | qm9/last.ckpt | icd4vay6 |
| QM9 (with H) | qm9_with_h/last.ckpt | zbtevof0 |
| FlyBrain | flybrain/last.ckpt | mxjhu6l3 |
| MEL2 | mel2/last.ckpt | n6yuk51z |
| Toy sequences | toy_sequence/alphabet_{5,10,...,160}.ckpt | noise_scale=5.0 sweep |

The 50-step Euler GuacaMol row of Tab. 1 was evaluated on run 17oqxx3h while
its last checkpoint was at epoch 59; `guacamol/epoch_59.ckpt` is that file.

All molecular checkpoints were trained with the precision schedule
`initial_precision=3.0`, `final_precision=12.0`; the synthetic-graph checkpoints
use `final_precision=20.0` (planar, SBM) and `25.0` (tree).  The repository
configs are pinned accordingly -- do not evaluate them with the generic
`ole_scheduler` default.

Download with:

```bash
python scripts/download_checkpoints.py
```