mkirkby commited on
Commit
06f37ef
·
verified ·
1 Parent(s): 5437505

Add model card

Browse files
Files changed (1) hide show
  1. README.md +74 -1
README.md CHANGED
@@ -1,3 +1,76 @@
1
  ---
2
  license: mit
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: mit
3
+ library_name: transformers
4
+ base_model: google/gemma-4-31B-it
5
+ pipeline_tag: text-classification
6
+ language:
7
+ - en
8
+ tags:
9
+ - decision-model
10
+ - classification
11
+ - calibration
12
+ - routing
13
+ - guardrails
14
+ - vision
15
+ - multimodal
16
+ - baseten
17
+ ---
18
+
19
+ # bison-1
20
+
21
+ 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.
22
+ Provide the model a state (text, optionally with images) and a set of questions and it will return a calibrated probability for every option,
23
+ answering all questions from **one shared-prefix prefill**, without generating text.
24
+
25
+ For faster, cheaper decisions, see the mixture-of-experts sibling [groundhog-1](https://huggingface.co/baselabs/groundhog-1).
26
+
27
+ ## 🗒️ Model Details
28
+
29
+ | Model | Parameters | Description |
30
+ |---|---|---|
31
+ | [Gemma-4-31B](https://huggingface.co/google/gemma-4-31B-it) | 31B | General-purpose base model |
32
+ | **[bison-1](https://huggingface.co/baselabs/bison-1)** | 31B | Post-trained for single-pass, calibrated decisions |
33
+ | [groundhog-1](https://huggingface.co/baselabs/groundhog-1) | 26B (about 4B active) | Mixture-of-experts sibling, fastest decisions |
34
+
35
+ - **Inputs**: text and images
36
+ - **Training**: full fine-tune of the language model on text and image decisions; the vision encoder, norms and embeddings are frozen
37
+ - **Attention**: block-bidirectional mask on full-attention layers; sliding-window layers stay causal.
38
+ - **Max sequence length**: 131,072 tokens
39
+
40
+ ## 🏃 How to use
41
+
42
+ A request carries one `state` and any number of `questions`. Every question is answered in the same call.
43
+
44
+ ```python
45
+ questions = {
46
+ "refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"},
47
+ "team": {
48
+ "type": "choice",
49
+ "instructions": "Which team should handle this?",
50
+ "criteria": {"billing": "Charges, refunds, invoices",
51
+ "technical": "App or site faults",
52
+ "fraud": "Suspected unauthorised use"},
53
+ },
54
+ "urgency": {
55
+ "type": "score",
56
+ "instructions": "How urgent is this?",
57
+ "criteria": ["Can wait", "Today", "Blocking the customer now"],
58
+ },
59
+ }
60
+ request = {"state": "I was charged twice this month, please refund one of them.", "questions": questions}
61
+ ```
62
+
63
+ | `type` | `criteria` | answer fields |
64
+ |---|---|---|
65
+ | `noul`: yes or no | optional | `noul`: P(yes) |
66
+ | `choice`: one of named options | `{name: description}` | `choice`, `probabilities` |
67
+ | `score`: ordered levels | list of levels, lowest first | `score`, `probabilities` |
68
+
69
+ Requests longer than the trained length limits are refused, not truncated.
70
+
71
+ ## 🔧 Training
72
+
73
+ - **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.
74
+ - **Objective**: cross-entropy plus a Brier calibration term (λ = 1), with a trained "I don't know" option.
75
+ - **Recipe**: 1 epoch (3,747 steps), peak LR 5e-6, no weight decay, 3% warm-up, bf16, sequences up to 131,072 tokens.
76
+ - **Hardware**: 24 nodes × 8 B200 (192 GPUs).