bison-1 / README.md
mkirkby's picture
Add LICENSE and NOTICE, state licences on the card
bf8350b verified
|
Raw History Blame Contribute Delete
5.6 kB
---
license: other
license_name: mit-and-apache-2.0
license_link: LICENSE
library_name: transformers
base_model: google/gemma-4-31B-it
base_model_relation: finetune
pipeline_tag: text-classification
language:
- en
tags:
- decision-model
- classification
- calibration
- routing
- guardrails
- vision
- multimodal
- baseten
---
<div align="center">
<img
src="https://huggingface.co/baselabs/bison-1/resolve/main/assets/banner.png"
alt="Baseten bison-1 decision model"
style="width: 100%; max-width: 100%; height: auto; display: inline-block; margin-bottom: 0.5em; margin-top: 0.5em;"
/>
<div style="display: flex; justify-content: center; gap: 0.5em; margin-bottom: 1em;">
<a href="https://www.baseten.co/library/bison-1/"><strong>Try on Baseten</strong></a>
</div>
</div>
# bison-1
bison-1 is a 31B parameter **decision model** post-trained from [Gemma-4-31B](https://huggingface.co/google/gemma-4-31B-it) by Base Labs at Baseten.
Provide the model a state (text, optionally with images) and a set of questions and it will return a calibrated probability for every option,
answering all questions from **one shared-prefix prefill**, without generating text. On tested benchmarks:
- **#1 of 19 decision models on the Artificial Analysis Decision Index v2.1**: 76, ahead of d1.2-ultra (75) and mammoth-1 (71).
- **Best-calibrated model on that index**: Brier score 0.14.
For faster, cheaper decisions, see the mixture-of-experts sibling [groundhog-1](https://huggingface.co/baselabs/groundhog-1).
## 🗒️ Model Details
| Model | Parameters | Description |
|---|---|---|
| [Gemma-4-31B](https://huggingface.co/google/gemma-4-31B-it) | 31B | General-purpose base model |
| **[bison-1](https://huggingface.co/baselabs/bison-1)** | 31B | Post-trained for single-pass, calibrated decisions |
| [groundhog-1](https://huggingface.co/baselabs/groundhog-1) | 26B (about 4B active) | Mixture-of-experts sibling, fastest decisions |
- **Inputs**: text and images
- **Training**: full fine-tune of the language model on text and image decisions; the vision encoder, norms and embeddings are frozen
- **Attention**: block-bidirectional mask on full-attention layers; sliding-window layers stay causal.
- **Max sequence length**: 131,072 tokens
## 🏃 How to use
### Baseten Model APIs
bison-1 is available on [Baseten Model APIs](https://www.baseten.co/library/bison-1/) as `decision-models/baseten-bison-latest`, at $0.10 per 1M tokens. It uses the same API as mammoth-1 and elephant-1, and the endpoint accepts text and image inputs.
### Request format
A request carries one `state` and any number of `questions`. Every question is answered in the same call.
```python
questions = {
"refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"},
"team": {
"type": "choice",
"instructions": "Which team should handle this?",
"criteria": {"billing": "Charges, refunds, invoices",
"technical": "App or site faults",
"fraud": "Suspected unauthorised use"},
},
"urgency": {
"type": "score",
"instructions": "How urgent is this?",
"criteria": ["Can wait", "Today", "Blocking the customer now"],
},
}
request = {"state": "I was charged twice this month, please refund one of them.", "questions": questions}
```
| `type` | `criteria` | answer fields |
|---|---|---|
| `noul`: yes or no | optional | `noul`: P(yes) |
| `choice`: one of named options | `{name: description}` | `choice`, `probabilities` |
| `score`: ordered levels | list of levels, lowest first | `score`, `probabilities` |
Requests longer than the trained length limits are refused, not truncated.
## 📊 Performance
### Artificial Analysis Decision Index v2.1
Scored by Artificial Analysis, November 19, 2026.
| Model | Decision Index |
|---|---:|
| **bison-1** | **76** |
| d1.2-ultra | 75 |
| Drex Max v1.1 | 73 |
| SubQ Snap 1.2 | 73 |
| GLiDE v1.2 | 73 |
| mammoth-1 | 71 |
| Benchmark | bison-1 | Rank of 19 |
|---|---:|---:|
| AA-Paperclip | 73 | 2 |
| EvalSafe-O*NET-AA | 93 | 1 |
| When2Call | 98 | 1 |
| AA-Omniscience-Decision v1.1 | 60 | 11 |
| HLE-MultiChoice v1.1 | 51 | 1 |
| AA-Symmetry v1.1 | 89 | 1 |
### Long context (our run)
| Benchmark | bison-1 |
|---|---:|
| LongMemEval M, 128k tokens | 84.8 |
| LongMemEval S, 128k tokens | 87.1 |
## 🔧 Training
- **Data**: 6.71M decision examples (Mix 12): core decision skills, safety and abstention, judging and expert domains, document and long-context tasks, and public decision data, including public benchmark sets.
- **Objective**: cross-entropy plus a Brier calibration term (λ = 1), with a trained "I don't know" option. About one example in seven carries a probability target instead of a single right answer.
- **Recipe**: 1 epoch (3,747 steps), peak LR 5e-6, no weight decay, 3% warm-up, bf16, sequences up to 131,072 tokens.
- **Hardware**: 24 nodes × 8 B200 (192 GPUs).
## License
bison-1 is fine-tuned by Baseten from [Gemma 4 31B IT](https://huggingface.co/google/gemma-4-31B-it), which is licensed under the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0). Baseten's modifications are released under the MIT License. Use of the model is subject to both licenses; see [LICENSE](LICENSE) and [NOTICE](NOTICE).
## Attribution
Base model: Gemma 4 31B IT, made available by Google LLC, licensed under the Apache License 2.0. Baseten modified the weights by further training. Gemma is a trademark of Google LLC. This model is not affiliated with or endorsed by Google.