Models
SauerkrautLM-decision
A decision model from VAGO Solutions that answers typed questions about an input with calibrated probabilities instead of generated text.
Overview
SauerkrautLM-decision takes an input (the state) and one or more typed questions, and returns one structured answer per question. Instead of free text, it returns probabilities your code can threshold, route on or rank by. TODO: one or two sentences on what the model is built for and what it was trained on.
Yes/No answers
A single probability that the answer is "yes", ready to threshold.
Choice and Score
Pick one option from a set, or place the input on an ordered scale, each with a full probability distribution.
Several questions per call
All questions are evaluated against the same input in one call.
Specifications
| Property | Value |
|---|---|
| Parameters | TODO |
| Context length | TODO |
| Architecture | TODO |
| Inputs | TODO (text only, or text and images?) |
| Outputs | Probabilities per question (Yes/No, Choice, Score). TODO confirm |
| Precision | bf16 |
| Checkpoint size (bf16) | TODO |
| Minimum GPU memory | TODO |
| Hugging Face | VAGOsolutions/SauerkrautLM-decision (TODO confirm repo name) |
| License | TODO |
Quick Start
Install:
pip install -U transformers torch accelerateTODO: confirm the minimum transformers version.
Load the model and ask three questions about one input:
from transformers import AutoModel
model_id = "VAGOsolutions/SauerkrautLM-decision" # TODO confirm repo name
model = AutoModel.from_pretrained(
model_id,
trust_remote_code=True,
dtype="bfloat16",
device_map="auto",
)
state = "Ich wurde diesen Monat doppelt belastet, bitte erstatten Sie eine Zahlung."
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 website faults",
"account": "Login, profile or subscription changes",
},
},
"urgency": {
"type": "score",
"instructions": "How urgent is this?",
"criteria": ["Can wait", "Handle today", "Blocking the customer now"],
},
}
output = model.system_one(state, questions)
print(output)TODO: confirm that the remote code exposes system_one and these question type names.
Request format
| Field | Description |
|---|---|
state | The input to evaluate: plain text or a JSON-serialisable object. |
questions | Named questions. Each has a type (noul, choice or score), instructions, and for Choice and Score, criteria. |
images | TODO: only if image input is supported. |
Question types
Type noul. Returns one probability between 0 and 1 that the answer is "yes". Values near 0 or 1 are clear answers; values near 0.5 mean the model is unsure.
questions = {
"is_complaint": {
"type": "noul",
"instructions": "Is this message a complaint?",
}
}Answer shape (TODO confirm):
{ "is_complaint": { "type": "noul", "noul": "TODO" } }Type choice. Picks one option from a set you define and returns the probability of each option, plus a confidence value.
questions = {
"team": {
"type": "choice",
"instructions": "Which team should handle this?",
"criteria": {
"billing": "Charges, refunds, invoices",
"technical": "App or website faults",
},
}
}Answer shape (TODO confirm):
{
"team": {
"type": "choice",
"choice": "billing",
"probabilities": { "billing": "TODO", "technical": "TODO" },
"confidence": "TODO"
}
}Type score. Places the input on an ordered scale you define (level 0 upward) and returns the probability-weighted position plus the distribution per level.
questions = {
"urgency": {
"type": "score",
"instructions": "How urgent is this request?",
"criteria": ["Can wait", "Handle today", "Blocking the customer now"],
}
}Answer shape (TODO confirm):
{
"urgency": {
"type": "score",
"score": "TODO",
"probabilities": { "0": "TODO", "1": "TODO", "2": "TODO" },
"confidence": "TODO"
}
}Using the answer
Because answers are numbers, the decision logic stays in your code:
p = output["refund"]["noul"] # TODO confirm output structure
if p > 0.8:
action = "auto_refund"
elif p < 0.2:
action = "no_refund"
else:
action = "human_review"The 0.8 / 0.2 thresholds are an example. Tune them on your own data.
Deployment
SauerkrautLM-decision is published as a single bf16 checkpoint. No GGUF, MLX, W8A8 or other quantized variants are provided. Plan for at least TODO GB of GPU memory. For hosted access without your own GPU, see the SauerkrautLM-decision API.