# gnn

JAX checkpoint for the `GatV2` shared-trunk graph model with separate intention and outcome heads.

## Checkpoint

- Model class: `GatV2`
- Framework: `jax`
- Run timestamp: `20260430_103325`
- Parameters: `model_params.joblib`
- Hyperparameters: `config.json`
- Metrics: `metrics.json`

## Validation Metrics

```json
{
  "val/loss": 1.1955294609069824,
  "val/intention_top1_accuracy": 0.6223060488700867,
  "val/intention_top2_accuracy": 0.8227370381355286,
  "val/intention_top3_accuracy": 0.9057111740112305,
  "val/intention_mrr": 0.7693313956260681,
  "val/outcome_top1_accuracy": 0.5634428858757019,
  "val/outcome_top2_accuracy": 0.7412446141242981,
  "val/outcome_top3_accuracy": 0.8213900923728943,
  "val/outcome_mrr": 0.7087604999542236
}
```

## Loading

```python
import json
from pathlib import Path

import joblib

checkpoint_path = Path("path/to/checkpoint")

with open(checkpoint_path / "config.json") as f:
    config = json.load(f)

params = joblib.load(checkpoint_path / "model_params.joblib")
```