wazn-2b-v0.1

Wazn is a decision model: given a context, optional rules and an instruction, it returns a probability distribution over a set of labels you supply, each with an optional definition and few-shot examples. Nothing is generated. The label set is yours, per request, with no retraining: two labels or a thousand. Large sets (hundreds to 1,000+ classes) are handled by tournaments.

Beta. For trying the model out, not for production. Requests run one at a time, and the API may change between 0.x releases.

Code, full docs and examples: github.com/numid-ai/wazn-experimental.

How it works

1. Encode each candidate. The rules, context and instruction are followed by one candidate label (its definition and examples), closed by a </choice> tag. The frozen Qwen3.5-2B backbone, with LoRA adapters, reads this sequence, and the hidden state at </choice> is projected into a representation of that candidate.
2. Compare the candidates. A relational judge compares the candidate representations against each other, rather than scoring each label alone, and outputs a probability for each one.
3. Check that any label fits. A NONE gate estimates the probability that at least one candidate is valid. Below a threshold the answer is NONE; otherwise it is the most likely candidate.

Use

Python 3.11+. Install the wazn-experimental package in a virtual environment, then start a local server with this checkpoint:

python -m venv .venv && source .venv/bin/activate
pip install --pre "wazn-experimental[server]"
wazn-experimental serve --model numidlabs/wazn-2b-v0.1
# serving on http://127.0.0.1:8000

• [server] installs PyTorch (a large download). On Linux with an NVIDIA GPU, use "wazn-experimental[server,cuda]" for the fast kernels.
• Besides a server, Wazn.load() runs the model in process (e.g. in a notebook).

Then query it from Python:

from wazn_experimental import Client, Request, Instruction, Label

request = Request(
    context="My running shoes arrived in the wrong size. Can I swap them?",
    instructions=[Instruction(
        "Which team should handle this?",
        labels=[Label("returns", "Exchanges, refunds, wrong or damaged items"),
                Label("shipping", "Delivery status, delays, lost packages"),
                Label("billing", "Charges, invoices, payment problems")],
    )],
)
answer = Client().predict(request).answer
print(answer.choice, answer.confidence)  # e.g. returns 0.93
print(answer.probabilities)              # distribution over the labels, sums to 1
print(answer.none_probability)           # P(none of the labels applies)

For more than a few dozen labels, give the instruction a tournament=Tournament(group_size=10, top_k=2). See the README.

Speed

Config	Instructions × labels	Input tokens (context / instructions / labels)	Median	p90
Normal context, few labels	1 × 5	105 (60 / 13 / 32)	35 ms	37 ms
Short query, banking77 intents (snake_case names)¹	1 × 77 (tournament, groups of 10, top 2)	492 (17 / 12 / 463)	48 ms	50 ms
Normal context, many labels, all at once	1 × 219	1,052 (60 / 13 / 979)	121 ms	122 ms
Normal context, many labels	1 × 219 (tournament, groups of 10, top 2)	1,052 (60 / 13 / 979)	86 ms	87 ms
Normal context, very many labels (random names)	1 × 1,000 (tournament, groups of 10, top 2)	5,483 (60 / 13 / 5,410)	317 ms	319 ms
Long context (NDA, ContractNLI)	1 × 10	6,485 (6,317 / 14 / 154)	152 ms	153 ms
Normal context, many instructions (incident report)	10 × 4	1,621 (648 / 189 / 784)	154 ms	156 ms

On an NVIDIA RTX 5090, bfloat16, [server,cuda]. Server-side time over 10 requests sent one after another, after a warm-up request; no input was truncated. The first request after a start is slower, because it compiles the GPU kernels. Without [cuda], or on Apple Silicon or CPU, expect much slower requests; see the README.

¹ 50 different banking77 test queries, one request each, rather than one request repeated; tokens are medians (queries run 9 to 30 tokens).

Limitations

TODO: known weak spots, training data summary, evaluation results.

License

The head weights are released under Apache-2.0. The backbone keeps its own license; see Qwen/Qwen3.5-2B-Base.