mkirkby commited on
Commit
d6cdabf
·
verified ·
1 Parent(s): 7c0bed4

Update model card

Browse files
Files changed (1) hide show
  1. README.md +27 -1
README.md CHANGED
@@ -16,12 +16,21 @@ tags:
16
  - baseten
17
  ---
18
 
 
 
 
 
 
 
19
  # groundhog-1
20
 
21
  groundhog-1 is a 26B parameter mixture-of-experts **decision model** (about 4B active parameters per token) post-trained from [Gemma-4-26B-A4B](https://huggingface.co/google/gemma-4-26B-A4B-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 the highest accuracy, see the dense 31B sibling [bison-1](https://huggingface.co/baselabs/bison-1).
26
 
27
  ## 🗒️ Model Details
@@ -39,6 +48,12 @@ For the highest accuracy, see the dense 31B sibling [bison-1](https://huggingfac
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
@@ -68,9 +83,20 @@ request = {"state": "I was charged twice this month, please refund one of them."
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 (5,621 steps), peak LR 3.95e-6, no weight decay, 3% warm-up, bf16, sequences up to 131,072 tokens.
76
  - **Hardware**: 16 nodes × 8 B200 (128 GPUs).
 
16
  - baseten
17
  ---
18
 
19
+ <div align="center">
20
+ <div style="display: flex; justify-content: center; gap: 0.5em; margin-bottom: 1em;">
21
+ <a href="https://www.baseten.co/library/groundhog-1/"><strong>Try on Baseten</strong></a>
22
+ </div>
23
+ </div>
24
+
25
  # groundhog-1
26
 
27
  groundhog-1 is a 26B parameter mixture-of-experts **decision model** (about 4B active parameters per token) post-trained from [Gemma-4-26B-A4B](https://huggingface.co/google/gemma-4-26B-A4B-it) by Base Labs at Baseten.
28
  Provide the model a state (text, optionally with images) and a set of questions and it will return a calibrated probability for every option,
29
  answering all questions from **one shared-prefix prefill**, without generating text.
30
 
31
+ - **Long documents at a small-model price**: trained on sequences up to 131,072 tokens and priced like elephant-1.
32
+ - **Same training data and recipe family as bison-1**, the #1 decision model on the Artificial Analysis Decision Index v2.1.
33
+
34
  For the highest accuracy, see the dense 31B sibling [bison-1](https://huggingface.co/baselabs/bison-1).
35
 
36
  ## 🗒️ Model Details
 
48
 
49
  ## 🏃 How to use
50
 
51
+ ### Baseten Model APIs
52
+
53
+ groundhog-1 is available on [Baseten Model APIs](https://www.baseten.co/library/groundhog-1/) as `decision-models/baseten-groundhog-latest`, at $0.035 per 1M tokens. It uses the same API as mammoth-1 and elephant-1, and the endpoint accepts text and image inputs.
54
+
55
+ ### Request format
56
+
57
  A request carries one `state` and any number of `questions`. Every question is answered in the same call.
58
 
59
  ```python
 
83
 
84
  Requests longer than the trained length limits are refused, not truncated.
85
 
86
+ ## 📊 Performance
87
+
88
+ ### Long context (our run)
89
+
90
+ | Benchmark | groundhog-1 |
91
+ |---|---:|
92
+ | LongMemEval M, 128k tokens | 77.4 |
93
+ | LongMemEval S, 128k tokens | 79.3 |
94
+
95
+ groundhog-1 has not yet been scored on the Artificial Analysis Decision Index.
96
+
97
  ## 🔧 Training
98
 
99
  - **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.
100
+ - **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.
101
  - **Recipe**: 1 epoch (5,621 steps), peak LR 3.95e-6, no weight decay, 3% warm-up, bf16, sequences up to 131,072 tokens.
102
  - **Hardware**: 16 nodes × 8 B200 (128 GPUs).