Rated typed decisions
Give SCM a text and a question, and get a probability for every answer. Pick one of several options, answer yes or no, or place the text on a scale.
SCM (Sovereign Classification Model) is a family of open models built on Qwen3, at 1.7, 4 and 14 billion parameters. Version 1 (v1) is the current release: a 14-billion and a 4-billion model. The 1.7-billion v1 model did not pass our release check and is not published. This repository holds the public documents (also on GitHub: https://github.com/5ak3t/scm). It contains no code.
Why SCM
The best typed-decision models today are very good and very large. The hosted Jev service is the accuracy reference, and the strongest open models sit around 27 billion parameters. Each asks you to accept something. With Jev your text goes to someone else's servers, every call is billed, and you cannot run the model yourself. With the 27-billion-parameter open models you need serious hardware to run them at all. At the other end, very small specialised models are fast but give up accuracy, and some are trained on the benchmark's own data families, which makes their scores hard to compare.
SCM sits between the two and is built to be inspected. There are three sizes. The 14-billion model has about half the parameters of the 27-billion class, and the 4 and 1.7-billion models fit modest hardware. Everything about them is open: the model files, every training source with its licence, the sources we kept out and why, and an audit of how much of a public benchmark overlaps our training data, published next to the results. Options are given with each question, so a new workflow needs a new list of options, not a new training run.
The evidence is modest and we state it plainly. On held-out workflow groups, v1 accuracy climbs from 59.3% at 1.7 billion parameters to 69.9% at 4 billion and 82.8% at 14 billion, so size is a real lever (those groups come from families v1 also trained on, so this flatters v1). On the public S1MB English benchmark the 14-billion v1 model completes all 137 benchmarks and scores 55.5, with a full disclosure of training overlap. Part of that gain comes from training on the training splits of the datasets S1MB is built from, so it is not a zero-shot number (see below).
SCM is not the most accurate typed-decision model. On the same benchmark the 14-billion v1 model scores below the hosted Jev 1.13 reference on yes/no questions (60.5 against 64.6), score questions (39.2 against 46.9) and, narrowly, choice questions (66.7 against 67.2), and below the 27-billion class. It gives plain probabilities with no calibration or error guarantee, handles up to 255 options per question, and some training sources are non-commercial or share-alike. If accuracy is all that matters, use the leaders. If you want an open, self-hostable model whose data and evaluation you can check, and you are willing to test it on your own data first, that is what SCM is for.
Results
All figures are percentages (higher is better). The size table uses our own held-out workflow groups and one training run per size, so treat small gaps as noise. The S1MB table is an outside benchmark.
Size comparison (v1, held-out workflow groups)
| Model | Held-out accuracy | Released |
|---|---|---|
| scm-mid-4b-v1 | 69.9 | yes |
| scm-max-14b-v1 | 82.8 | yes |
| 1.7B v1 | 59.3 | no (did not pass the release check) |
S1MB English v1 (baseline-adjusted score, 0 to 100, all 137 benchmarks complete, zero failed decisions)
| Yes/no (59) | Choice (57) | Score (21) | Task average | Rank* | |
|---|---|---|---|---|---|
| scm-max-14b-v1 | 60.5 | 66.7 | 39.2 | 55.5 | 19 of 165 |
| scm-mid-4b-v1 | 45.9 | 38.2 | 6.3 | 30.2 | 94 of 165 |
| Jev 1.13 (reference) | 64.6 | 67.2 | 46.9 | 59.6 | 17 of 165 |
*Our own computation from the saved results, ordering by the leaderboard's Borda rule. The shared leaderboard may order differently. The v1 results are accepted and merged into the S1MB results dataset: 14B https://huggingface.co/datasets/hotchpotch/s1mb-result/discussions/38 and 4B https://huggingface.co/datasets/hotchpotch/s1mb-result/discussions/40.
Read the v1 numbers with this in mind. v1 was trained on the training splits of 65 dataset families that S1MB benchmarks are built from (79 declared "same source family" entries), after removing inputs that matched S1MB test inputs. An audit of the final training pool still finds 304 of 15,609 English test cases with an approximate match, mostly in templated families. These are candidates, not confirmed use of test inputs, and no exact use is known. Compare with models that did not train on these families with care. Details: docs/s1mb-submission.md.
Models (Hugging Face)
| Release | Base | Link |
|---|---|---|
| scm-max-14b-v1 | Qwen3-14B-Base | https://huggingface.co/5ak3t/scm-max-14b-v1 |
| scm-mid-4b-v1 | Qwen3-4B-Base | https://huggingface.co/5ak3t/scm-mid-4b-v1 |
Each repository holds the model weights (a small add-on to the Qwen3 base model), a model card, the training-data lists with source licences, and the prompt format. They are experimental research releases. There is no 1.7B v1 release: it was trained but did not pass our held-out release check (how v1 was built and checked).
Documents
- How v1 was built and checked: training data, release check, readout experiments that did not help.
- Results: charts and the S1MB summary.
- How to use the models: how to ask a question, how the answer is scored, option limits.
- Training data and licences: sources by group, recorded licences, synthetic tickets, exclusions.
- Intended use and limits.
- S1MB submission: configuration, training-overlap disclosure, results pull requests.
- Changelog.
Status
Selection-level results on open development data, mostly single-seed. The 14B v1 model is below the strongest S1MB entries (rank 19 of 165 by our computation, with the training-overlap caveat above). Read the intended-use page before relying on these weights.
Cite
@misc{scm2026,
title = {SCM: rated typed decisions},
author = {5ak3t},
year = {2026},
howpublished = {\url{https://github.com/5ak3t/scm}},
note = {Version 1}
}
A CITATION.cff is included.
Contact
Report issues at https://github.com/5ak3t/scm/issues (or the Community tab of the Hugging Face repositories).
Licence
The documents and figures in this repository are CC BY 4.0 (see LICENSE). The model weights are separate and have their own terms; see data and licences.