Apertus-ParlTopics-8B: Apertus 1.5 8B as a decision model for parliamentary topics

Apertus-ParlTopics-8B is a LoRA adapter and a yes/no head for the text part of Apertus 1.5 8B. Given the text of a Swiss parliamentary affair and one or more topic cards (a name, a definition and a list of what belongs to the topic), it returns one probability per card in a single pass. It is the decision model behind the stored corpus results of Parlamentsthemen.

Hugging Face: https://huggingface.co/merabro/Apertus-ParlTopics-8B

The Hugging Face repository holds the adapter, the head, the adapter and training configurations and the training metrics. It does not include the Apertus tokenizer or weights: accept the terms of swiss-ai/Apertus-v1.5-8B, convert its text part with python -m parltopics.convert_text, and copy its tokenizer files into the folder before serving.

Files and checksums

File	Size	SHA-256
adapter_model.safetensors	159 MB	61df7d434bc3662f6dadea5eb63ea213a617b343e1d06fcaad04b39894d975a0
head.pt	8.4 MB	cb93496bcf2ada57f00e133944b4b4eac5e8daa3b6d9908d616d3ac8eda0747f

The checksums were computed on 19 October 2026 on the local copy and on the copy that the running server loads (the GPU server's run folder on the GPU server); both match. The adapter configuration, the training configuration and the training metrics are in this folder (adapter_config.json, training_config.json, training_metrics.json).

Model

Item	Value
Base model	swiss-ai/Apertus-v1.5-8B, text part only
Conversion	python -m parltopics.convert_text keeps the language-model weights of the multimodal checkpoint and writes them as an ApertusForCausalLM checkpoint; the base weights are not changed
Training and serving code	Kev (https://github.com/jaredpalmer/kev, Apache-2.0) at upstream commit 1f04cfa, with third_party/kev/apertus.patch (one file, kev/model.py: Apertus special tokens and the beginning-of-sequence token, and LoRA targets limited to the projections that exist in Apertus)
Adapter	LoRA rank 16, alpha 32, dropout 0.05, on q_proj, k_proj, v_proj, o_proj, up_proj, down_proj (PEFT 0.21.0)
Head	Kev's pointer head (projection dimension 256) on the model's hidden states, trained from scratch; output: probability that the affair is about the card
Precision	bf16 weights and training

Kev is an independent open implementation of the decision-model idea behind TypeSafe's System One models.

Training data

Item	Value
Records in the training file	6,753 (one affair text with several yes/no questions per record), covering 6,553 distinct federal affairs
Records used	6,728 with 31,837 questions; 25 records were dropped because they exceeded the training context (1,024 state tokens, 1,664 per question branch, 2,688 packed)
Affairs	Federal affairs (Curia Vista) dated before 2024; no affair from 2024 on, and none from the development or test samples
Questions	23 of the 28 Curia Vista subject areas, and the 400 most frequent Curia Vista descriptors (thesaurus keywords, assigned 1996 to mid-2014)
Never training questions	The subject areas Energie, Landwirtschaft, Kultur, Steuer and Migration; digitalisation; every cantonal and city scheme; APS. Related subjects appear among the descriptors
Labels	The Parliamentary Services' assignments in Curia Vista, via the OpenParlData label folder; not an independent gold standard
Built with	parltopics.build_train and parltopics.build_extra

Training configuration

From training_config.json and training_metrics.json:

Setting	Value
Epochs	1 (841 optimizer steps)
Learning rate	5e-5 for adapter and head (head_lr 0 means the same rate), one-cycle schedule with 10% warm-up
Batch	1 record, gradient accumulation 8
Weight decay	0.01
Label smoothing	0.1
Distractor options	p_none 0.1, p_none_distract 0.12, p_distract 0.15, perm_frac 0.3
Maximum state	1,024 tokens
Hardware and time	one RTX A6000 (48 GB), 6,491 s (1.8 h), peak GPU memory 42.8 GB, 7.06 million forward tokens
Training seed	0 (Kev's --seed)
Sampling seed	20261018, for the selection of training affairs, the development and test samples and all evaluation samples

Intended use

• Assigning topics defined as cards to Swiss parliamentary affairs from federal, cantonal and municipal parliaments, in German, French and Italian: standard schemes (Curia Vista areas, APS chapters and topics, cantonal schemes) and ad-hoc topics written by an operator.
• Ranking and shortlisting for people who read the results, for example a weekly monitoring list. Decisions are suggestions; each stored result carries the card hash, the model revision and, when calibrated, a probability.
• Not intended for decisions about persons, for inferring the political position of an affair or its authors, or for use without human review where a wrong topic has consequences.

Evaluation

All figures are from docs/README_overview.md. Test affairs are from 2024 to 2026, with thresholds fitted on earlier affairs; the same period was also used to choose between training runs.

Evaluation	Apertus-ParlTopics-8B	For comparison
Digitalisation, 3,000 affairs, all levels (weighted to the PolitMonitor evaluation universe)	AP 0.46, ROC-AUC 0.97; precision 0.36, recall 0.77, F1 0.49 at the default threshold	untrained Apertus 8B AP 0.43; untrained Apertus 70B 0.41; embedding classifier trained on the labels 0.47
Digitalisation, monitoring threshold	recall 0.97 at precision 0.18; 29 of the 30 DVS-listed affairs
Calibration, digitalisation	calibration error 0.022
Curia Vista areas, 1,000 federal affairs	macro AP 0.71, micro F1 0.70	untrained Apertus 8B 0.51; classifiers trained on the labels 0.73 to 0.74
The five withheld areas	macro AP 0.83	untrained Apertus 8B 0.63
St. Gallen scheme, 868 cantonal affairs (never a training question)	macro AP 0.70	untrained Apertus 8B 0.57
APS chapters, 1,000 federal affairs, against the Curia Vista mapping	macro AP 0.74, micro F1 0.63
APS topics, 500 federal affairs, against the frontier-model judge	precision 0.54, recall 0.71, F1 0.61
30 descriptors outside training, holdout up to 2014	macro AP 0.70, macro F1 0.65, micro F1 0.60
Same 871 federal affairs in German, French, Italian	Curia Vista macro AP 0.71, 0.70, 0.70
Party groups, 1,000 federal affairs	Curia Vista micro F1 between 0.66 and 0.74; no aggregate over- or under-assignment established
Throughput, one RTX A6000	4.9 affairs/s for one card; 0.83 affairs/s for 28 cards

The frontier-model judge is Claude Opus 5.5 (Claude Sonnet 5.5 for the digitalisation samples), used only as a second reader in the evaluation.

Limitations

• The test period 2024 to 2026 was also used to choose between runs; a later untouched period should confirm the figures.
• Title-only affairs (about 57,600 in the corpus) are often missed at a single threshold; the gap between French- and German-speaking parliaments at the default threshold comes mostly from missing text.
• Individual decisions differ between languages for the same federal affair: of the decisions where at least one language says yes, only 66% say yes in all three.
• On narrow topics, hosted frontier models find more (higher recall), for example on data-protection affairs within digitalisation.
• APS quality rests on the Curia Vista mapping and the judge, because no APS-labelled affairs exist; topic thresholds use a shared default.
• New cards have no calibrated probability and no validated threshold until labelled affairs are available.
• Scores vary slightly with the request. Scoring the same cards again in the same grouping changed scores by at most 0.13 on the logit scale on two affairs, without changing a decision. Scoring a card alone instead of together with the other corpus cards in one request can move a score enough to cross a threshold: for one title-only affair (334249), the stored digitalisation score is 1.76 and the score of the card alone 1.86, on either side of the threshold 1.81.
• No Romansh test set exists. Party bias was checked on federal affairs only.

Licence

• The adapter and the head (this model): Apache-2.0, copyright 2026 Invincibles.
• The base model Apertus 1.5 8B: Apache-2.0, with the Apertus Acceptable Use Policy (https://github.com/swiss-ai/apertus-legal/blob/main/apertus_1.5/USAGE_POLICY.pdf), which users accept when they download the gated base model and which applies to its use with this adapter.
• Kev and the patch to it: Apache-2.0.
• Training labels: Curia Vista data of the Parliamentary Services of the Federal Assembly, via OpenParlData.ch (CC BY 4.0, "Source: OpenParlData.ch").

How to serve it

Kev's server loads the base model, the adapter and the head and answers POST /v1/systemone. On our GPU server:

scripts/gpu/serve_model.sh 1 8350

which runs, inside a Kev checkout at 1f04cfa with apertus.patch applied,

CUDA_VISIBLE_DEVICES=1 HF_HUB_OFFLINE=1 KEV_TRUNCATE_STATES=1 python -m kev.serve --run <folder with this adapter> --port 8350

and waits until GET /v1/models answers. The folder holds adapter_config.json, adapter_model.safetensors, head.pt, the tokenizer files and training_config.json, whose base entry points to the converted text model. Requests are built with parltopics.systemone_multi.build_request(text, {card_id: card}, compact=True); the answer answers[card_id]["noul"] is the probability, and logit of it is the score that the calibration files in models/ map to a calibrated probability and a decision. Set PARLTOPICS_DECISION_URL to the server's address so that the API and the nightly job use it.