AirDecision-4B-v1.2
an updated model for structured decisions
2 min readupdated
AirDecision-4B-v1.2 is a standalone text classification and decision-making system built on unsloth/Qwen3.5-4B. It handles option scoring, binary probabilities, and ordinal expectations in a single forward pass with zero generated output tokens. Compared to the previous release, it expands the training dataset using O*NET documents and improves phrasing symmetry. The model targets developers and researchers working on automated evaluation and testing on benchmarks like JevBench.
Key features
- Distribution-based training: uses soft-distribution cross-entropy for accurate probability calibration.
- Data synthesis: generates training examples from specifications with double-blind quality checks.
- Phrasing invariance: stays consistent despite changes in option order and question wording (negation symmetry).
- Zero generation overhead: scores options in a single pass without autoregressive text generation.
- Isolated runtime: ships with a ready-to-use local HTTP server featuring strict context-length checks that return errors when limits are exceeded.
What you can do
- Spin up a local server to score decisions via JevBench integration scripts.
- Integrate the model into Python apps to automatically classify incoming tickets or tasks against specific criteria.
- Run reproducible evaluations of business logic and choice options inside air-gapped containers with no internet access.
Example
import sys
sys.path.insert(0, "./AirDecision-4B-v1.2")
from inference import DecisionEngine
engine = DecisionEngine("./AirDecision-4B-v1.2", device="cuda", dtype="bfloat16", max_length=8192)
response = engine.predict({
"state": {"ticket": "Invoice charged twice for subscription renewal", "tier": "pro"},
"questions": {
"route": {
"type": "choice",
"instructions": "Select the handling queue.",
"criteria": {
"billing": "Payment, invoice, or refund issues",
"technical": "Product bugs or outages",
},
},
},
})
print(response["answers"])
Read more
DetailsBenchmarks: averages 91.38% accuracy across 7 test sets, and scores 77.48% (86/111) on JevBench Hard.
Hardware requirements: GPU supporting bfloat16 dtype; runs on standard PyTorch architecture with SDPA.
Limitations: maximum context length of 8192 tokens, up to 62 answer choices per question, and zero output tokens.










Comments