Pusula

Pusula

An open Turkish decision engine that answers typed questions about a message, such as routing, urgency, human handoff or prompt injection, with a probability for every option and no text generation.

Made by erenextlabs 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: Pusula is a Turkish decision engine. It starts from the vllm-sr/d3 base model, called Decision 3.0, and adds training on Turkish data. For each message you ask typed questions, and it returns a probability for every possible answer in one pass.

Questions you can ask include which queue a request belongs in, how pressing it is, whether to pass it to a person and whether it smuggles in an injected prompt. The repository is a LoRA adapter plus code, about 0.9 GB, and the current version is v2.0.

Who it suits: Teams that handle Turkish messages and want fast, structured triage, plus developers who build agents that need to check inputs. A GPU environment that can run the 26B base model is needed.

Industry | Who uses it | How they use it | Example
Customer support | Support teams | Route and rate incoming Turkish messages | Deciding if a message needs a human agent
Finance | Service desks | Flag urgent or fraud-related requests early | A message about unknown card withdrawals
AI agents | Agent developers | Check inputs for hidden instructions and personal data | Screening an email before an agent reads it
Moderation | Community teams | Detect insults and hate speech in Turkish text | Tagging abusive messages for review

Anyone who works in languages other than Turkish, or who wants free-form text replies, can skip it. It answers questions and does not generate text.

Where to use it:

• Python, through the pusula package, which applies the code readout the adapter needs.
• Windows, macOS or Linux with Python, for downloading the files.
• Support pipelines and agent systems that need a structured decision per message.

Use case | Who | What they get back
Message routing | Support teams | A probability for each routing option
Urgency scoring | Service desks | A score across levels from routine to critical
Safety checks | Agent developers | Probabilities for prompt injection or personal data in a message

How to set it up: You need Python, the pusula package and the base model files.

• Python with huggingface_hub
• The pusula package with its d3 extra, installed from the project's own repository
• The base vllm-sr/d3 model, a 26B Apache 2.0 model
• A GPU setup that can run it

1. Install the pusula package with its d3 extra, as the project's own instructions say.
2. Download the base model without its assets folder.
3. Download the adapter into a local folder.

hf download vllm-sr/d3 --local-dir d3 --exclude "assets/*"
hf download erenextlabs/pusula --local-dir pusula

How to use it: Load the engine with the base model and the adapter, then call decide with a message and the questions you want answered. Loading the adapter with generic PEFT is not enough, because this package applies the code readout and attention hook.

from pusula.inference import D3Engine

engine = D3Engine.load("d3", adapter="pusula")
answers = engine.decide(
    "Kartımdan bilgim dışında üç çekim yapıldı, hemen engelleyin!",
    {
        "temsilci": {"type": "noul", "instructions": "Bir insan temsilciye aktarılmalı mı?"},
        "aciliyet": {"type": "score", "instructions": "Ne kadar acil?", "criteria": ["Rutin", "Yakın", "Yüksek", "Kritik"]},
    },
)

Tool | What it does
D3Engine.load | Loads the base model and the Pusula adapter
engine.decide | Answers your typed questions about one message
score question | Rates a message across levels you list, such as routine to critical

Example: A customer writes in Turkish that three withdrawals were made from their card without consent and asks for it to be blocked. You ask whether a human should take over and how urgent it is, from routine to critical. The engine returns a probability for every option.

For newcomers and experts: If you are new, think of Pusula as a triage helper that reads a Turkish message and answers your multiple-choice questions about it. Try one message with two questions, as in the sample.

For experts, the adapter is LoRA with rank 32, applied on the 26B d3 base through a code readout. Custom questions and the missing-information template are answered by the base model with the adapter off. The card reports 90.6 on held-out real customer messages. On the public benchmark v2.0 scores 92.0, a little under v1.0 at 92.6, mainly on urgency, human handoff and domain.

A Turkish message arrives
You define typed questions with options
The engine loads d3 with the Pusula adapter
One forward pass scores every option
You read the probabilities per question

Title: Pusula 2.0 accuracy in percent
Real customer messages: 90.6%
Long customer complaints: 88.3%
Turkish Decision Benchmark: 92.0%
Gold test set: 92.1%
Custom questions in unseen sectors: 93.5%
MASSIVE tr-TR intent: 85.2%
OffensEval 2020 Turkish: 88.3%

Good to know: The points below cover limits and costs.

• Language: Turkish.
• Training data: synthetic messages for this project, plus customer messages licensed from a commercial partner for the first adapter. Those real messages are not published.
• License: Apache 2.0.
• Cost to run: no API key or paid service is named. The 26B base model needs a capable GPU.
• Data it reaches: it reads the messages you pass in. Downloads come from Hugging Face.
• Not for: loading the adapter with plain PEFT, languages other than Turkish, or free-text generation.

On Hugging Face, 13 November 2026

• Task: text classification
• Library: pusula
• Formats: Safetensors
• Downloads in the last 30 days: 23
• Likes: 2
• License: apache-2.0
• Languages: tr
• Based on: vllm-sr/d3
• Tags: pusula, safetensors, decision-model, turkish, routing, lora, d3, text-classification

Links

• Model: https://huggingface.co/erenextlabs/pusula
• Files: https://huggingface.co/erenextlabs/pusula/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download erenextlabs/pusula

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