Darwin-27B-ZTC

Darwin-27B-ZTC

Reads a situation and a set of typed questions, then returns a probability for every option in one forward pass, with no text generated. For builders adding calibrated decision steps to workflows.

Made by FINAL-Bench on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: A model of the Darwin family that makes decisions without generating any tokens. You give it a piece of state and typed questions, and it returns a probability for each option of each question. It needs one forward pass per question and writes no text. Each option may come with a short description, which the model reads along with the question.

Who it suits: Developers who want a model to choose between listed options and report how sure it is, instead of writing an answer in prose.

Industry | Who uses it | How they use it | Example
Customer support | Support operations teams | Pick the next action for a ticket from a fixed list | A billing ticket scored across refund, escalate or close
Workflow automation | Automation builders | Add a yes or no check to a pipeline | A probability that a request needs approval
Review and compliance | Reviewers | Score a case against an ordered rubric | A probability for each level of the rubric
AI engineering | Agent builders | Use it as a fast decision step beside a chat model | Route work to the right tool with calibrated odds

You can skip it if you need free text, chat or summaries, or if your GPU cannot hold about 54 GB of weights.

Where to use it: It runs as Python code on a GPU machine, using the inference code that ships inside the repository.

• Python with torch, transformers with Qwen3.5 support, safetensors and pillow
• A GPU that can hold roughly 54 GB of BF16 model weights
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Ticket routing | Support teams | A probability for each action on the list
Yes or no checks | Workflow builders | The odds of yes against no
Rubric scoring | Reviewers | A probability for each level of an ordered rubric

How to set it up: Install the packages, download the repository, and put its folder on your Python path so the shipped autojev code can load. The repository also carries the weights and the readout head.

• Python
• torch, transformers with Qwen3.5 support, safetensors and pillow
• A GPU with room for the BF16 weights

1. Install the four Python packages.
2. Download the repository snapshot with the Hugging Face tools.
3. Add the downloaded folder to your Python path.
4. Create a DecisionModel from that folder.

pip install -U huggingface_hub
hf download FINAL-Bench/Darwin-27B-ZTC

How to use it: Build a row with a state and a question. The question has a type, instructions and criteria, where each criterion is an option with its own description. The call returns one probability per option, in the order of the criteria.

import sys
from huggingface_hub import snapshot_download

path = snapshot_download("FINAL-Bench/Darwin-27B-ZTC")
sys.path.insert(0, path)
from autojev.model import DecisionModel

model = DecisionModel(checkpoint=path)
row = {
    "state": {"ticket": "Customer was charged twice for the same order."},
    "question": {
        "type": "choice",
        "instructions": "What should support do?",
        "criteria": {"refund": "Refund the duplicate charge.", "escalate": "Send to billing.", "close": "No action."},
    },
}
probabilities = model.predict([row])[0]  # one probability per option, in criteria order

Tool | What it does
noul | A yes or no question type
choice | A question type that picks one of N labels
score | A question type that rates against an ordered rubric
DecisionModel | Loads the checkpoint folder
model.predict | Returns one probability per option
readout.safetensors | The decision readout head
decision_config.json | Holds option codes, the calibration temperature and provenance

Example: Describe a ticket about a customer billed twice, ask what support should do, and list refund, escalate and close with a short description each. You get back three probabilities that add up across the options, with no sentence generated.

For newcomers and experts: Most language models write text. This one only scores the options you hand it, so the answer is a set of numbers that show how likely each option is. A first small try is the card's ticket example, then change the options and watch the numbers move.

For experts, the backbone is Darwin-27B-RSI trained in full as a decision engine. Each option gets a short code inside the prompt. A small readout head scores the options from the model's last hidden state, and a temperature fitted on data calibrates the softmax that turns scores into probabilities. The training data came from public benchmarks, using their train and dev parts and never their test parts. The trade is a closed set of options and a large GPU in exchange for calibrated odds with no token generation.

Describe the state and the question
List options with short codes
Run one forward pass
Readout head scores each option
Softmax gives the probabilities

Good to know: The card reports results on the Typed Decisions test split, which has 400 cases and 2,000 decisions, with an overall accuracy of 0.743.

Title: Accuracy by question type
Yes or no (noul): 0.845
Choice: 0.732
Score (rubric): 0.675
Overall: 0.743

• Zero-shot: the model never saw the benchmark's train split, its workflows or its question schemas.
• Run noise: two runs of the test set gave accuracy 0.741 and 0.743, which the card puts down to batching.
• Errors: all 2,000 decisions were answered with no errors.
• Shipped code: the repository includes MIT-licensed autojev code for decision models.
• License: apache-2.0.
• Cost to run: none named, no API key, account or paid service, though it needs a large GPU.
• Data it reaches: the state and questions you give it, plus the files downloaded from Hugging Face.
• Not for: free-text generation, since it generates no tokens.

On Hugging Face, 13 November 2026

• Task: text classification
• Library: transformers
• Parameters: 25.6B
• Weights: about 51.2 GB in BF16; about 12.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 326
• Likes: 40
• License: apache-2.0
• Languages: en
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5_text, feature-extraction, darwin, zero-token-classifier, decision-engine, typed-decisions, system-one, calibration, probabilistic-classification, text-classification

Links

• Model: https://huggingface.co/FINAL-Bench/Darwin-27B-ZTC
• Files: https://huggingface.co/FINAL-Bench/Darwin-27B-ZTC/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download FINAL-Bench/Darwin-27B-ZTC

Then follow the usage in its model card. Nothing in this kit runs by itself.