QEV1.1-9B

English · 한국어 · 日本語

Inference-only release · merged LoRA · BF16

Overview

QEV multilingual typed-decision DLP model with Japanese-focused augmentation. It evaluates is_dlp and direct is_* type questions while retaining the QEV choice, noul, and score interface. It returns option probabilities, not generated text.

Base: Qwen/Qwen3.5-9B-Base · revision: 68c46c4b3498877f3ef123c856ecfde50c39f404 · format: BF16

Training data

The project training corpus combines publicly available general-purpose data with internal corporate DLP data and multilingual augmentation. The released repository does not include raw training, calibration, validation, or test rows, labels, generated examples, optimizer state, or adapter files. Dataset names, internal paths, sample counts, and generation rules are intentionally not disclosed.

Self-contained package

This inference-only release contains the complete merged backbone, tokenizer, FP32 pointer head, calibration metadata, and custom Transformers loader. It does not require a private or Dev repository, a separate adapter, or an external base-model download. Training scripts, adapter files, optimizer state, conversion tools, and datasets are not distributed. Publishing full weights does not technically prevent further fine-tuning; applicable license terms still govern use.

Usage

from transformers import AutoModel
model = AutoModel.from_pretrained("QEV1.1-9B", trust_remote_code=True, device="cuda:0")
print(model.decide("This is a public operational notice.", {"dlp": {"type": "noul", "instructions": "Does this contain sensitive data?", "criteria": {"false": "No", "true": "Yes"}}}))

Use the complete repository with AutoModel and decide(). Generic text-generation pipelines and hosted inference are not validated. The LoRA update is merged into the BF16 backbone; the pointer head remains FP32.

Evaluation (historical source reference)

These are recorded direct type-question results for the unmerged BF16 source checkpoint, not a new full benchmark of this merged/8bit package. validation.json records serialized reload and nine-question smoke checks.

Metric	Result
Overall Macro-F1 (biometric excluded)	97.72%
EN Macro-F1	96.69%
KO Macro-F1	98.29%
JA Macro-F1	98.46%
Evaluation records	6,281

is_biometric has zero positive rows and is excluded from macro summaries. See EVALUATION.md, evaluation.json, and evaluation/ for details. Raw training/evaluation rows, optimizer state, and adapters are not distributed.

Limits and intended use

The package is self-contained: backbone, tokenizer, pointer head, custom loader, serving example, evaluation metadata, and quantization/training guidance are included. MANIFEST.sha256, provenance.json, and validation.json record the contents. The interface performs document-level decisions and does not extract finding spans or counts. Validated hardware is NVIDIA H200/CUDA; custom code requires trust_remote_code=True.

License and attribution

The QueryPie-trained modifications, merged LoRA update, pointer head, calibration metadata, and custom code use QueryPie AI Research License v1.0. Foundation weights and tokenizer assets retain their upstream licenses. See LICENSE, LICENSES/Apache-2.0.txt, THIRD_PARTY_LICENSES.md, and NOTICE for scope and attribution. Training data is not relicensed or distributed.

Personal non-commercial research and experimentation, and non-commercial academic research and evaluation are permitted. Commercial use requires prior written permission from QueryPie or a separate commercial license agreement. For commercial use, contact QueryPie. Business operations, product development and commercially motivated proofs of concept are not covered by the free research permission. Third-party foundation components retain their original licenses.