Skip to main content

Neuralk, the predictive AI company

Neuralk OSS

Neuralk-AI open-source models.

Seldon

Seldon is a tabular in-context-learning foundation model: fit keeps the training table as context, predict_proba classifies new rows against it in one forward pass, with no gradient step.

from neuralk_oss import SeldonClassifier

clf = SeldonClassifier.from_pretrained("seldon-mini")  # or "seldon"
clf.fit(X_train, y_train)  # pandas DataFrame (or array) and labels
proba = clf.predict_proba(X_test)  # (n_rows, n_classes), columns in clf.classes_ order

A model's name loads it with the settings it is meant to run with, from its packaged config (see Configs): seldon with 8 ensemble members, seldon-mini with 4, half of them reading categories in a shuffled order. Keyword arguments override them, e.g. from_pretrained("seldon", device="cpu").

Checkpoints

model checkpoint task parameters
seldon seldon-classifier classification, up to 10 classes per model call 213.5M
seldon-mini seldon-mini-classifier classification, up to 10 classes per model call 27.4M

SeldonClassifier.from_pretrained takes the model name and loads its classification checkpoint; Seldon.from_pretrained takes the checkpoint name.

A checkpoint is a directory holding config.json and model.safetensors. from_pretrained takes:

  • a checkpoint name of the table above (or, for SeldonClassifier, a model name): downloaded from its Hugging Face repository, Neuralk-AI/seldon or Neuralk-AI/seldon-mini, where the checkpoint is a directory of that name, on the first call, then read from the Hugging Face cache (~/.cache/huggingface, or HF_HOME);
  • a local directory, read in place: from_pretrained("path/to/seldon-mini-classifier");
  • any Hugging Face repository, with the checkpoint's directory and a revision: from_pretrained("org/repo", subfolder="seldon-mini-classifier", revision="v1").

The repositories are private for now: log in once with an account that can read them (hf auth login), or pass token=... to from_pretrained.

Column types are inferred (categorical columns are embedded as such, NaNs are kept for the model's own imputation), each of the n_estimators ensemble members (8 by default) sees its own preprocessing, and their logits are averaged. Inference runs in float16 autocast on CUDA by default (amp=False for float32). Tables with more classes than the model's head (10) go through error-correcting output codes.

Preprocessing

The preprocessing and ensembling come from structured-data-models (sdm, Apache-2.0), pinned in pyproject.toml. SeldonRecipe builds its recipe, and takes its parameters as keyword arguments:

from neuralk_oss import SeldonClassifier, SeldonRecipe

recipe = SeldonRecipe(transforms=("identity", "rank_gaussian"), softmax_temperature=0.9)
clf = SeldonClassifier.from_pretrained("seldon-mini", recipe=recipe)
parameter default
transforms ("identity", "power", "robust", "rank_gaussian") numerical transforms the members take in turn, one each
shuffle_categories False the members read categorical codes in a random order per column, instead of sorted; for some members only, see below
members "all" the members this recipe preprocesses: see below
flip_sign True each member negates a random half of its numerical columns
feature_shuffle "latin" column permutation per member: "latin", "random" or None
class_shift True each member shifts the class codes by its own offset
average_logits True average the members' logits, else their probabilities
softmax_temperature 1.0 divisor of the logits before the softmax
max_features 500 columns the model reads, cut after the shuffle

The rest (clipping bounds, rank-Gaussian knots, category counts) is in SeldonRecipe's docstring. Beyond parameters, each step is an overridable method:

import sdm.processing as sp
from neuralk_oss import SeldonClassifier, SeldonRecipe


class MyRecipe(SeldonRecipe):
    max_features = 200

    def numerical_transforms(self):
        return [sp.Identity(), sp.RankGaussian(max_knots=self.rank_gaussian_knots)]


clf = SeldonClassifier.from_pretrained("seldon-mini", recipe=MyRecipe())

Members can take different recipes: a list of recipes, each with the members it preprocesses, which together cover every member once. Each group runs as its own ensemble, and the groups' outputs are averaged in log space, weighted by their sizes:

recipes = [
    SeldonRecipe(members=[0, 1, 2, 3], shuffle_categories=False),
    SeldonRecipe(members=[4, 5, 6, 7], shuffle_categories=True),
]
clf = SeldonClassifier.from_pretrained("seldon-mini", n_estimators=8, recipe=recipes)

Configs

For full control over a run, beyond from_pretrained's defaults, a YAML config sets the checkpoint, the classifier and its recipe in one file: checkpoint, any keyword argument of from_pretrained, and recipe, SeldonRecipe's keyword arguments, or a list of them, one per group of members. Each released model's settings are a packaged config of its name, which from_pretrained loads: seldon and seldon-mini.

# seldon.yaml
checkpoint: seldon
n_estimators: 8
max_context_size: 200000
random_state: 0
amp: true
recipe:
  - members: [0, 1, 2, 3]
    shuffle_categories: false
  - members: [4, 5, 6, 7]
    shuffle_categories: true
clf = SeldonClassifier.from_config("seldon")  # the same as from_pretrained("seldon")
clf = SeldonClassifier.from_config(
    "my_run.yaml", device="cpu"
)  # keyword arguments override the file

template.yaml, packaged next to them, lists every argument a config can set, at its default, with what it does: copy it and keep the keys you change. Keys a config leaves out take the model's own settings for seldon and seldon-mini, the constructor's defaults otherwise. To change n_estimators of a model whose config splits its members into groups, set recipe too, so that its members match.

The model alone

from neuralk_oss import Seldon

model = Seldon.from_pretrained("seldon-mini-classifier")
context = model.encode_context(X_context, y_context, categorical_mask)  # (B, T, H), (B, T), (B, H)
logits = model.predict_query(X_query, context, categorical_mask)  # (B, T_query, 10)

Contributing

See CONTRIBUTING.md: layout, development and versions.

License

Apache 2.0.

Metadata

Release files for neuralk-oss 0.2.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for neuralk-oss 0.2.0
File Size Uploaded Provenance Actions
neuralk_oss-0.2.0.tar.gz 916.1 kB With provenance Details for neuralk_oss-0.2.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for neuralk-oss 0.2.0
File Interpreter ABI Platform Actions
neuralk_oss-0.2.0-py3-none-any.whl Python 3 none any Details for neuralk_oss-0.2.0-py3-none-any.whl

Total release size: 955.1 kB

Release files / neuralk_oss-0.2.0.tar.gz

Download URL neuralk_oss-0.2.0.tar.gz
Size 916.1 kB
Tags Source
SHA-256 checksum
How to use checksums
6bd09e907c9c030faa8740736db083666f12dafd58dd01b0d9ee726755897589
BLAKE2b-256 checksum
How to use checksums
80f099d54870e98ac6c2c5f275a9d15747908d363cb9e26bb22a82fd53d30582
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

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.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 2, 2026.

Transparency log

Release files / neuralk_oss-0.2.0-py3-none-any.whl

Download URL neuralk_oss-0.2.0-py3-none-any.whl
Size 39.1 kB
Tags Python 3
SHA-256 checksum
How to use checksums
799f128806e487d4505e1f1085ce910b2e2c609233f51801c439c6288623f993
BLAKE2b-256 checksum
How to use checksums
3547cdc0395fd8f190d411871ebfb7192f989afea0628227d66c854354f9a3a1
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

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.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 2, 2026.

Transparency log

Security for neuralk-oss 0.2.0

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.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for neuralk_oss-0.2.0.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 2, 2026.

Transparency log

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.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page