typical-ai
Typed, direct decisions from a pretrained language model. State + question + a label set you define at call time go in; a probability distribution over exactly those labels, plus an explicit abstain, comes out. One forward pass, nothing generated, nothing to parse.
A self-contained decision head for OzLabs/typical-small-preview, OzLabs/typical-small, and
OzLabs/typical-medium. No training-repo dependencies (no bench/train/data/
metrics) -- just torch, transformers, huggingface_hub, safetensors, numpy.
Install
pip install typical-ai
The import name is typical_ai, not typical: PyPI's typical is an unrelated, established
package, so taking that import name would break anyone who has both installed.
from typical_ai import Typical
m = Typical.from_pretrained("OzLabs/typical-small") # or OzLabs/typical-medium
m.choice(state, "What does the customer want?", ["refund", "replacement", "repair"])
m.noul(state, "Is the order still under warranty?")
m.score(state, "How urgent is this ticket?", ["0", "1", "2", "3"])
Models and their cards, including the measured trade-offs of each release, are at https://huggingface.co/OzLabs. Source and the full experimental record: https://github.com/GuyNachshon/typical.
(Or just copy the typical/ directory next to your code -- it's a plain Python package,
no build step.)
Usage
from typical_ai import Typical
m = Typical.from_pretrained("OzLabs/typical-small", device="auto") # or "typical-small-preview" / "typical-medium"
# K-way choice over a fixed label set -> {label: p, ...} + p_null
m.choice(state, "What does the customer want?", ["refund", "replacement", "repair"])
# Yes/no -> P(yes)
m.noul(state, "Is the order still under warranty?")
# Ordinal levels -> {level: p, ...} + p_null + expected (E[index] under the candidates)
m.score(state, "How urgent is this ticket?", ["0", "1", "2", "3"])
# Full JevBench-style decide() -- mirrors pcdm_jev.decider.PCDMDecider.decide exactly,
# including the runtime block (latency, p_null, truncation flags).
probs, runtime = m.decide(
state,
{"type": "choice", "instructions": "What does the customer want?",
"criteria": {"refund": "money back", "replacement": "a new item shipped"}},
["refund", "replacement"],
)
state is either a string or a JSON-serialisable dict (dicts are dumped with
json.dumps before tokenising, same as the original decider).
See example.py for a runnable end-to-end script.
What's inside
typical/ is a trimmed, numerically-identical port of this repo's native-readout serving
path (pcdm_jev.decider.PCDMDecider(mode="native")):
typical/backbone.py-- frozen Qwen3 trunk truncated attap_layer, LoRA on its toplora_layersblocks (pcdm/encode.py'sBackbone, minus the training-only FeatureCache/EmbedEncoder/forward() machinery the serving path never touches).typical/native.py--NativeHead(factored null, the per-row Bernoulli "noul" route, theletters/tags/letters_nonull/query_onlyrenderers) andnative_kv_decide(encode the state once into a KV cache, score every query's suffix against it).typical/core.py--Typical.from_pretrained(downloadsbest.pt, readscheckpoint["args"]for backbone id /tap_layer/lora_r/lora_layers/nc_head/nc_render/null/noul_head/score_headto rebuild the exact backbone+head shape, then loads the LoRA + tower weights) plus thechoice/noul/score/decideAPI andquery_text/to_labels(verbatim frompcdm_jev/decider.py).
Not ported: nc_head values n2/n2n3 (Qwen3-Embedding candidate vectors) -- neither
shipped checkpoint trains with them; native_kv_decide raises NotImplementedError if a
future checkpoint needs that path. z-score calibration (--zscore) needs no special
handling -- its buffers live in the checkpoint's tower state_dict like any other weight.
Verification
inference/test_parity.py (gated behind RUN_SLOW=1, downloads the real 1.7B checkpoint)
runs 6 items through both the original PCDMDecider(mode="native") and Typical on the
same device and checkpoint, and asserts identical probabilities (max abs diff was 0.0
against guychuk/pcdm-runs/typical-small/best.pt on both CPU and MPS), plus the noul
reversed-label control (["no","yes"] vs ["yes","no"] give the same P(yes) through the
Bernoulli head) and a check that typical-small-preview's noul_head="choice" checkpoint
still answers noul() correctly through the plain K-way path.
RUN_SLOW=1 python inference/test_parity.py
Metadata
Release files for typical-ai 0.1.1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| typical_ai-0.1.1.tar.gz | 18.9 kB | Without provenance | Details for typical_ai-0.1.1.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| typical_ai-0.1.1-py3-none-any.whl | Python 3 | none | any | Details for typical_ai-0.1.1-py3-none-any.whl |
Total release size: 37.9 kB
Release files / typical_ai-0.1.1.tar.gz
| Download URL | typical_ai-0.1.1.tar.gz |
|---|---|
| Size | 18.9 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
e3051798d9c39035bb54c252419dd906b06696e21548a177224859067f96bf5a
|
|
BLAKE2b-256 checksum How to use checksums |
228e92adf5c780a169fa7ba94ffa4ed44db3ee4e0035a6e59e655745c528c423
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
uv/0.7.9
|
Release files / typical_ai-0.1.1-py3-none-any.whl
| Download URL | typical_ai-0.1.1-py3-none-any.whl |
|---|---|
| Size | 19.0 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
134f2bf015835c8dc1a2c75d62fa98596d4feafa2e20f5f44f2e89ba4e9c31e6
|
|
BLAKE2b-256 checksum How to use checksums |
2f7c1ec58fc9907f24a14a1d3e78e40e4d93c8cc98dda0f11947fe321c1d2e15
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
uv/0.7.9
|
Security for typical-ai 0.1.1
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.