Datasets:
Bongard-mini: general, zero-shot, open weights (accuracy 0.594, KL 0.256, Brier 0.132)
We scored AgentBull/bongard-mini on test (400 cases, 2,000 decisions) with full distributions. Mode: general, zero-shot, open weights.
| Model | Kind | Accuracy ↑ | KL from gold ↓ | Brier ↓ | ECE ↓ | p50 latency | Price / 1M input |
|---|---|---|---|---|---|---|---|
| Bongard-mini (T5Gemma 2 4B-4B + judgment head) | general, zero-shot, open weights | 0.594 | 0.256 | 0.132 | 0.067 | 225 ms‡ | open weights |
‡ Per case (one request with all of its questions), on one RTX 4090 in BF16, one request at a time, client and server on the same machine. Not comparable with hosted latency.
Other metrics: soft accuracy 0.518, macro F1 0.420, TV 0.257, score MAE 0.460, within one level 0.931. Accuracy by type: noul 0.763, choice 0.550, score 0.499.
It clears the Prior on accuracy, KL and Brier. Among the general models on the board, only meraGPT Decider 1 has a lower KL and a lower Brier.
What it is. A typed-decision model that generates no text. The backbone is Google's T5Gemma 2 4B-4B encoder-decoder, fully post-trained, with a judgment head. The encoder reads the state once, each question is a separate decoder sequence, and the head gives one logit per option. The distribution is the softmax of those logits divided by a per-type temperature from temperatures.json, which ships with the weights.
Zero-shot. The model never saw the train or test split of this dataset or its four workflows. Its training data includes multi-question workflow decisions in the same typed format on 30 other workflows.
Reproduce. Weights at revision c81c5b883a5983669be84be98658eac940489613, server at github.com/AgentBull/bongard commit 3e6e603:
pip install "bongard @ git+https://github.com/AgentBull/bongard@3e6e6030a953dfe8d30d93aca2d0ab92b87b4a5a"
hf download AgentBull/bongard-mini --revision c81c5b883a5983669be84be98658eac940489613 --local-dir bongard-mini
bongard serve --checkpoint bongard-mini --device cuda --temperatures bongard-mini/temperatures.json --port 8000
python benchmarks/typed_decisions.py --endpoint http://127.0.0.1:8000 --output predictions.json # in the bongard repository
The server speaks POST /v1/systemone, so the client you used for Jev and Jeff should work against it unchanged. Each case is one request with its state and all five questions.
Predictions. Every probability of every decision: typed-decisions/predictions.json.