Beko2210/statim-decide-multilingual-base-pii

Statim Decide Multilingual Base: PII adapter

Licence notice (2026-10-03). The base model this adapter requires (multilingual 0.7.0) was trained partly on data that is non-commercial, ShareAlike or under an unknown licence. From 2026-10-03 the base and this adapter are offered only under PolyForm Noncommercial 1.0.0; the Small Business and Free Trial licences and the commercial licence do not apply to them. A replacement trained only on cleared data will follow. Details: DATA_LICENSES.md.

A LoRA adapter that improves the PII decisions of Beko2210/statim-decide-multilingual-base 0.7.0. Checked with Statim 0.8.3: the published f32 file loads it merged at load, the q8_0 file as runtime LoRA. LoRA adapters need Statim 0.8.0 or later.

Quick start

hf download Beko2210/statim-decide-multilingual-base statim-decide-multilingual-base-q8_0.gguf --local-dir models
hf download Beko2210/statim-decide-multilingual-base-pii statim-decide-multilingual-base-pii.lora.gguf --local-dir models
statim serve -m multilingual=models/statim-decide-multilingual-base-q8_0.gguf --adapter multilingual:pii=models/statim-decide-multilingual-base-pii.lora.gguf --port 8080

curl -s localhost:8080/v1/systemone -d '{"state": {"text": "Hi, this is Anna Schmidt. Call me back at +49 170 1234567."}, "questions": {"pii": {"type": "noul", "instructions": "Does the text contain a phone number?"}}, "adapter": "pii"}'
curl -s localhost:8080/v1/systemone -d '{"state": {"text": "Hi, this is Anna Schmidt. Call me back at +49 170 1234567."}, "questions": {"pii": {"type": "noul", "instructions": "Does the text contain a phone number?"}}, "adapter": "auto"}'

Python (Python SDK 0.8.3 or later):

from statim import Client
client = Client("http://127.0.0.1:8080")
client.decide({"text": "Hi, this is Anna Schmidt. Call me back at +49 170 1234567."}, {"pii": {"type": "noul", "instructions": "Does the text contain a phone number?"}}, adapter="pii")

Results (experiment's f32 run)

Each cell uses 150 items (seed 20260927).

Language	Base	Adapter	Change (points)	Verdict	Qwen3-8B zero-shot
ar	0.8467	0.9067	+6.00	within noise	0.893
de	0.8400	0.8667	+2.67	within noise	0.893
en	0.8933	0.9267	+3.34	within noise	0.887
es	0.8533	0.8933	+4.00	within noise	0.867
fr	0.8933	0.9400	+4.67	within noise	0.880
it	0.8400	0.9200	+8.00	gain (2 SE)	0.873
ja	0.8467	0.9267	+8.00	gain (2 SE)	0.893
nl	0.7600	0.8667	+10.67	gain (2 SE)	0.800
ru	0.9067	0.9467	+4.00	within noise	0.947
sv	0.8333	0.8733	+4.00	within noise	0.800
zh	0.9000	0.9467	+4.67	within noise	0.920
Mean	0.8558	0.9103	+5.46	—	0.878

The pooled family change is +5.46 points with 2 SE = 2.22 points (gain).

A single cell of 150 items rarely clears 2 SE on its own; the decision uses the pooled family.

Decision rule: promote when the category family gains more than 2 standard errors, pooled by rows or by suites, and nothing regresses. A regression is a pooled drop beyond 2 standard errors under either pooling, or a drop in one language cell that stays significant after Holm-Bonferroni (family-wise 5 %).

Statim is trained on this category, while Qwen3-8B runs zero-shot.

Checked on the published files

Weights	Adapter mode	Base mean	Adapter mean	Change (points)	Cell for cell as in the experiment
f32	merged at load	0.8558	0.9103	+5.46	yes
q8_0	runtime LoRA	0.8558	0.9109	+5.52	— (experiment: f32)

Training

Source	Rows	Licence
E3-JSI/synthetic-multi-pii-ner-v1 (default)	2,971	MIT
Wismut/nym-pii-multilingual-data (default)	6,200	MIT
gretelai/gretel-pii-masking-en-v1 (default)	6,200	Apache-2.0
gretelai/synthetic_pii_finance_multilingual (default)	6,200	Apache-2.0
nvidia/Nemotron-PII (default)	6,200	CC-BY-4.0
urchade/synthetic-pii-ner-mistral-v1 (data.json)	6,200	Apache-2.0

LoRA rank 16, alpha 32.0, dropout 0.05; target modules Wqkv, Wo, Wi; 88 wrapped modules and 3,379,200 trainable parameters.

• Items: 33,571 train, 400 dev.
• Updates: 1,172.
• Dev accuracy: 0.8375 before, 0.8725 after.
• Time: 1631.7 seconds; peak memory: 2,332 MB.

Provenance

• Adapter GGUF SHA-256: 2991a33b5d9db4f5b679081885faab1f84289d6eae7660830447e790bd168b29
• PEFT safetensors SHA-256: f8e7e674ab868b312ff32f403552daa436e7437411e83fa0ece60187431a2d6f
• Training checkpoint model.safetensors SHA-256: c44425f14ac9d55508f73a6f371e4e2802ed59e646287c3abbec10b653a19840
• Base fingerprint: e7a8fa743b4920850167b587226e339d651be0244fcc6a9e24fff8ed5073bae5
• Training mixture SHA-256: b8a87e85f3bf72b61509555aba9148358760d3329ecf37a36cebd94f9358cb03
• Source registry SHA-256: 309ce8ec262886b3bfaa529934151ae0c1879595c26301e4269da276d95cbfcd

Experiment commands (paths relative to the Statim repository):

• train: '.venv-train/bin/python' 'tools/finetune/train_lora.py' 'models/laya-multilingual-v9' --mixture 'data/mixture-v8.jsonl.gz' --category pii --registry 'tools/finetune/sources/v6-keep.json' --out 'models/lora-exp1/pii' --device cuda --epochs 2
• convert: '.venv/bin/python' 'tools/convert_lora.py' 'models/lora-exp1/pii' -o 'models/lora-exp1/pii.lora.gguf' --base 'models/laya-multilingual-v9-f32.gguf' --category pii --name pii
• serve: 'build-vk/statim' serve -m 'multilingual=models/laya-multilingual-v9-f32.gguf' --adapter 'multilingual:pii=models/lora-exp1/pii.lora.gguf' --device vulkan --threads 16 --port 8098 --no-access-log --inference-timeout 600
• eval_base: '.venv/bin/python' 'bench/eval_categories.py' --strict --url http://127.0.0.1:8098 --model multilingual --suites pii --n 150 --seed 20260927 --exclude-mixture 'data/mixture-v8.jsonl.gz' --out 'models/lora-exp1/pii.base.jsonl'
• eval_adapter: '.venv/bin/python' 'bench/eval_categories.py' --strict --url http://127.0.0.1:8098 --model multilingual --suites pii --n 150 --seed 20260927 --exclude-mixture 'data/mixture-v8.jsonl.gz' --adapter pii --out 'models/lora-exp1/pii.adapter.jsonl'

Protocol and experiment: docs/ADAPTERS.md. Full reproduction instructions: REPRODUCE.md.

Intended use and limits

• This adapter only helps its category; route requests with "pii" or "auto".
• Bound to Beko2210/statim-decide-multilingual-base 0.7.0 by the base fingerprint e7a8fa743b492085… (statim.lora.base_fingerprint, SHA-256 over the checkpoint's norm and bias tensors); Statim refuses the adapter on a base whose fingerprint differs.
• Languages outside the evaluated list are untested.
• Each language cell has 150 items.
• Do not automate decisions about people without human review.

Licence

The weights may be used under any one of: PolyForm Noncommercial 1.0.0, PolyForm Small Business 1.0.0 (free commercial use below 100 people and 1 M USD revenue), PolyForm Free Trial 1.0.0 (any company, fewer than 32 days), or a Statim commercial licence (COMMERCIAL.md). Texts in LICENSE-MODEL.md. The Statim engine is Apache-2.0.

Training data attribution is listed source by source above, with the row count and licence read from the experiment registry.