Pusula

An open-source AI decision engine for Turkish. Pusula is Decision 3.0 (vllm-sr/d3) fine-tuned on Turkish decision data. For each message, typed questions (for example, how a request should be routed, how urgent it is, whether a human should take over, or whether it contains a prompt injection) are answered with a probability for every option in a single forward pass, without text generation.

Doğru karar, doğru yön.

Version	v1.0 (18 October 2026); the preview weights of 17 October 2026 remain available as revision v1.0-onizleme
Base model	vllm-sr/d3 (26B, Apache-2.0)
This repository	LoRA adapter (rank 32) + code readout, ~0.9 GB
Language	Turkish
Training data	Specially curated Turkish data: synthetic messages generated for this project and human-written customer messages licensed from a commercial partner
Interface	Same request and answer format as d3's /v1/systemone

Results

Accuracy (%). Evaluation labels were assigned blind and independently by two annotators (Claude Opus 5.5 and Claude Sonnet 5.5); disagreements were adjudicated, and all labels follow Pusula's written labeling policy. The real customer messages were labeled blind by Claude under the same policy. Evaluation labels are used exclusively for evaluation and are never used as training labels.

Evaluation	Questions	Decision 3.0	Pusula
Real customer messages (500 held-out, licensed)	4,500	77.4	85.6
Long customer complaints (500 held-out, licensed)	4,500	81.4	88.0
Turkish Decision Benchmark (2,194 messages)	19,861	88.5	92.6
Gold test set (759 held-out messages)	7,015	85.5	90.4
Your own questions (497 messages × 3 ad-hoc questions)	1,491	94.8	94.2
MASSIVE tr-TR, 60-way intent, zero-shot (human-written Turkish)	2,974	85.3	85.2
OffensEval 2020 Turkish, offensive language (real tweets)	3,528	88.4	88.3

Per question:

Question	Gold set: Decision 3.0	Gold set: Pusula	Benchmark: Decision 3.0	Benchmark: Pusula
Urgency (4 levels)	53.2	78.9	76.5	86.6
Human handoff	81.5	89.2	77.1	88.6
Domain	82.1	86.8	89.9	94.0
Emotion	85.0	88.3	71.3	81.0
Security risk	94.2	98.0	97.5	99.2
Action required	88.9	90.5	89.2	90.0
Personal data	96.4	96.3	93.6	94.8
Intent	96.8	97.0	97.9	98.0

The two real evaluation sets (customer messages, and long, detailed customer complaints) were licensed from a commercial partner, are held out from training and are not published. Urgency is partly subjective: the two annotators agree on 84.4% of gold-set messages, which bounds the attainable score. MASSIVE and OffensEval are not included in the training data.

Usage

Pusula is loaded with the pusula package. Generic PEFT loading of the adapter is not sufficient, since the code readout and the non-causal attention hook of the d3 prompt contract are applied by this package.

pip install "pusula[d3] @ git+https://github.com/hsnsahinhsn/pusula"
hf download vllm-sr/d3 --local-dir d3 --exclude "assets/*"
hf download hsnsahinhsn/pusula --local-dir pusula

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"]},
    },
)

An HTTP server (pusula-serve), a built-in Turkish taxonomy (20 domains, 86 subdomains, 299 intents) and the benchmark are provided in the Pusula repository. A smaller model with the same interface is available as Pusula Nano (149M, CPU-capable), which can be tried on Google Colab without installation:

Hardware requirements: approximately 64 GB of GPU memory in BF16, or approximately 24 GB when loaded in 4-bit.

Training

• Data: about 13,500 synthetic Turkish messages across 20 domains, written by Qwen3.8-27B (Apache-2.0) from fixed targets (intent, urgency, emotion, style, persona, length); every question was answered again blind by a separate judge pass and by the base model, and only labels on which they agree are kept. TabiBERT was used to remove duplicates and likely label errors and to build hard intent distractors. Plus about 7,400 real messages licensed from a commercial partner and anonymized prior to use (about 3,500 customer messages and about 4,000 long customer complaints), labeled by Qwen3.8-27B under the written labeling policy; about 2,800 synthetic messages written by Qwen3.8-27B in the style of real customer messages; and about 8,900 messages with ad-hoc questions of many options, answered by the base model. Messages of the evaluation sets are excluded from training.
• Labeling policy: all training labels follow the written labeling policy, which was applied by Qwen3.8-27B with explicit decision rules (urgency, human handoff and opinion-only reviews).
• Method: a BF16 LoRA (rank 32, all attention and MLP projections) and the code readout were trained, continuing from an earlier Pusula adapter, for 225 optimizer steps over 24,000 questions, with a doubled share for ad-hoc questions. The checkpoint with the highest validation accuracy (step 100) is published.

Limitations

• Urgency is partly subjective. Where urgency drives costly actions, the probabilities should be used rather than only the top level.
• Emotion is the weakest built-in question on the benchmark (81.0). A bias toward "angry" over "disappointed" in critical reviews, observed in the preview release (76.3), was reduced by a revised labeling rule.
• Most of the training data is synthetic. Real traffic may differ in style and topic mix; evaluation on application-specific data is recommended before deployment.

Contributors and citation

Pusula is developed by the following contributors:

Name	Role	GitHub	Hugging Face
Hasan Şahin	Project lead	hsnsahinhsn	hsnsahinhsn
Ahmet Sayman	Contributor	asayman	asaymantr

Source code: github.com/hsnsahinhsn/pusula.

@software{pusula2026,
  title   = {Pusula: The First Full-Featured Turkish Decision Model},
  author  = {Şahin, Hasan and Sayman, Ahmet},
  year    = {2026},
  version = {1.0},
  url     = {https://github.com/hsnsahinhsn/pusula},
  license = {Apache-2.0}
}

License and attribution

Released under Apache-2.0. Pusula is built on vllm-sr/d3 (vLLM Semantic Router Team, Apache-2.0), which is itself built on Qwen3.8-27B (Apache-2.0). Qwen3.8-27B was used for data generation and labeling, and boun-tabilab/TabiBERT (Apache-2.0) for data curation. External evaluations: Amazon MASSIVE (CC BY 4.0) and OffensEval 2020 Turkish (Çöltekin, 2020; CC BY). See NOTICE.