DANTE v1
DANTE (Decisions As Next-Token Estimates) is a class-leading 4B decision model. You give it evidence, a question and the options; it returns a probability for every option from one forward pass. It generates no text: the answer is the model's next-token distribution over the option letters, so a decision costs one prefill and the probabilities are usable as confidence.
Try it in the playground: huggingface.co/spaces/elliottshort/dante-v1
It handles three kinds of question:
| Kind | Options | Example |
|---|---|---|
| Pick one | 2 to 26 named options | Which team should own this ticket? |
| Yes / no | yes, no | Did the agent stay within the user's request? |
| Score | ordered levels, lowest first | How urgent is this ticket, 0 to 4? |
llama.cpp builds: dante-v1-GGUF, including a Q4_K_M for 16 GB Macs.
Results
Accuracy at one option order. DANTE v1 was scored as the Q8_0 GGUF through llama.cpp and the untuned base in BF16 with PyTorch. The development sets were used to pick checkpoints, so read them as in-distribution numbers.
| Set | Items | Qwen3.5-4B, untuned | DANTE v1 |
|---|---|---|---|
| Development (mixed decision tasks) | 1,179 | 64.3 | 84.5 |
| Families (state tracking, tool-call guardrails, lead scoring) | 400 | 47.2 | 88.8 |
| Puzzles (temporal, financial and rule-following state) | 475 | 45.1 | 84.4 |
| Written decision scenarios | 697 | 93.0 | 97.7 |
Usage
Tested with transformers 5.17.
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import hf_hub_download
import importlib.util
path = hf_hub_download("elliottshort/dante-v1", "dante.py")
spec = importlib.util.spec_from_file_location("dante", path)
dante = importlib.util.module_from_spec(spec)
spec.loader.exec_module(dante)
tokenizer = AutoTokenizer.from_pretrained("elliottshort/dante-v1")
model = AutoModelForCausalLM.from_pretrained("elliottshort/dante-v1", dtype=torch.bfloat16)
model.to("cuda" if torch.cuda.is_available() else "cpu")
ticket = {"ticket": "The invoice total is double the price we were quoted.", "customer_tier": "enterprise"}
print(dante.decide(model, tokenizer, ticket, "Which team should own this ticket?", ["billing", "shipping", "technical support"]))
print(dante.decide(model, tokenizer, ticket, "Is the customer reporting a billing problem?", kind="noul"))
print(dante.decide(model, tokenizer, ticket, "How urgent is this ticket?",
{"0": "can wait a week", "1": "within two days", "2": "today"}, kind="score"))
dante.py builds the prompt the model was trained on and reads the letter probabilities:
<|im_start|>user
Answer with one option letter.
{evidence}
Q: {question}
A. {option}
B. {option}
Answer:<|im_end|>
<|im_start|>assistant
<think>
</think>
- Evidence and question may be strings or JSON; JSON is written compactly.
- Write an option as its name, or pass a dict to give each name a description (
name: description). - Yes / no questions list
yesthenno. Score levels go lowest first. - The temperatures in
dante.py(pick one 1.05, yes/no 0.35, score 0.5) calibrate one option order.viewsaverages more option orders to cancel position bias, at proportional cost.
Training
LoRA (rank 32) on Qwen/Qwen3.5-4B at revision 851bf6e8, merged into the weights you download here.
- 64k decision items: public decision sets, items labelled with rationales by a 27B teacher, written decision scenarios, and programmatically generated items with exact labels (state tracking, temporal and financial arithmetic, policy checks, agent traces). The loss is cross-entropy on the option letter, with a ranked probability loss on score questions.
- A second pass of two epochs on 6,000 training items the first pass got wrong or answered with less than 0.6 confidence, mixed with 6,000 it already solved; the adapter is merged at scale 0.8.
Limitations
- English only, and tuned for short decisions over supplied evidence, not open-ended generation or chat.
- Fine-grained subjective ratings, such as helpfulness or similarity scores, are a weak spot.
- Long agent traces and invoice matching remain the weakest in-distribution families.
- The probabilities are calibrated on our sets; recalibrate the temperatures on your own data if calibration matters.
Credits
Qwen3.5-4B by the Qwen team (Apache 2.0). Training data includes eikos-decisions by Caio Vicentino (CC BY 4.0), plumb-decisions by crh225, ARC (CC BY-SA 4.0), GSM8K (MIT), CommonsenseQA, StrategyQA and LogiQA.
Citation
@misc{short2026dantev1,
title = {{DANTE} v1: Decisions As Next-Token Estimates},
author = {Short, Elliott},
year = {2026},
month = oct,
url = {https://huggingface.co/elliottshort/dante-v1}
}