decider-0.8b: typed decisions with calibrated probabilities in one forward pass

The small sibling of decider-2b: a language model that does not generate text. It reads a state (a string or any JSON value) and a set of typed questions and returns a probability distribution for every question from a single forward pass: Choice (2-255 options, optionally described), Score (2-10 described levels), Noul (probability of yes). No decoding, no parsing, no output outside the options you defined. Same code, same wire format (POST /v1/systemone, TypeSafe Jev's format), same training recipe as the 2B: one epoch of scripts/train.sh full from Qwen/Qwen3.5-0.8B-Base over the full mixture (1.47M examples, 455M tokens, 4.5 h on one GH200). Code: https://github.com/Mapika/decider

# pip install git+https://github.com/Mapika/decider
from decider.infer import Decider
d = Decider("Mapika/decider-0.8b")
d.system_one(
    {"ticket": "I was charged twice for order A-104. Please refund the duplicate."},
    {"team": {"type": "choice", "instructions": "Which team should handle this?",
              "criteria": {"billing": "Charges, invoices, refunds", "technical": "Bugs, outages", "other": None}},
     "refund_requested": {"type": "noul", "instructions": "Does the customer ask for a refund?"},
     "frustration": {"type": "score", "instructions": "How frustrated is the customer?", "criteria": ["calm", "frustrated", "very frustrated"]}})

How it compares with the 2B

Same 93 public tasks, same protocol, one temperature fitted on in-task data (it came out at 1.03 for both: the recipe calibrates at every size). "Held-out" means no example of that dataset was trained on.

	decider-0.8b	decider-2b (same single-run recipe)
in-task accuracy / ECE, 69 tasks	0.776 / 0.032	0.809 / 0.030
held-out accuracy / ECE, 24 tasks	0.707 / 0.096	0.739 / 0.086
schema-first layout (the cacheable one), in-task / held-out accuracy	0.770 / 0.699	0.790 / 0.707
teacher-written custom questions, held-out domains: noul / choice / score	0.94 / 0.95 / 0.81	0.98 / 0.97 / 0.83
terse-bucket routing, held-out domains: generic / specific / catch-all	0.87 / 0.93 / 0.84	0.91 / 0.94 / 0.92
JSON state, one of 16 / 64 records named by path (64 with indices written in)	0.58 / 0.52 (0.62)	0.59 / 0.53 (0.63)
all 64 / 50 / 70 / 219 labels at once: HWU64, TREC-fine, DBpedia L2, L3 (held-out)	0.81 / 0.63 / 0.66 / 0.83	0.85 / 0.76 / 0.72 / 0.86
QuALITY, whole article (5-8k tokens)	0.63	0.68
isolated Score levels vs listwise (teacher-written, held-out domains)	0.83 vs 0.82, fits sum to 1.02	0.84 vs 0.84
hand-written battery: generic option right / catch-all right	0.95 / 0.95	0.90 / 0.85

What the smaller model gives up is knowledge, not the decision format: the largest drops are TruthfulQA (0.41 vs 0.55), OpenBookQA (0.67 vs 0.81), HellaSwag (0.76 vs 0.88) and ARC (0.74 vs 0.86), and wide label sets that need fine distinctions (TREC-fine). Routing, classification, yes/no judgments and JSON lookups on short states are within one to four points of the 2B. It is a weaker player: Pong and Breakout stay at the scripted teacher's level, CliffWalking fails (it walks off the cliff), held-out Freeway scores 0. The regression set runs about 1.5x faster than on the 2B; bf16 weights are 1.5 GB.

Limitations

Those of decider-2b, more so: a small model without reasoning; English only; rules written into the question ("fill if empty, otherwise skip") are not followed reliably, so state the decision as a plain question with described options; knowledge-heavy multiple choice is close to the base model; calibration is measured on public datasets and teacher-labelled probes, not on your traffic. The teacher-written training data comes from Qwen3.5-27B and carries its biases.