sinatras/torug-safety-bert-balanced-q8

Torug Turkish safety classifier — balanced BERT q8

Local Turkish prompt classifier exported from Başak's selected model_balanced checkpoint for the Torug 1.1 release candidate. This repository is separate from the September sinatras/torug-safety-bert-q8 model, which remains unchanged.

Artifact and labels

• Architecture: 12-layer BERT, hidden size 768, three output labels.
• Maximum sequence length: 512 tokens.
• Source safetensors SHA-256: ff25e26c8ae782cac235e9dcd1551e4dfe0dae259ddf348fef9a8c447dce35ae.
• ONNX opset 17; dynamic batch/sequence axes; per-channel dynamic QUInt8, matching the earlier release's quantization convention. Load with dtype: "q8".
• Weights: 111,846,943 bytes; SHA-256 5609f42b1dccef09dcf4dd6365d87f38ffabff2c407dfcea950190f9c31f09cf.
• Complete inference files total approximately 113 MB. See artifact-manifest.json for exact sizes and hashes. Pin a repository commit in production.

The checkpoint's generic labels are retained unchanged. Torug interprets LABEL_0 as safe, LABEL_1 as warn and LABEL_2 as block, consistent with its earlier named-label checkpoint and targeted benign/harmful/borderline probes. The original training label encoder and split manifest were not supplied. Do not infer calibrated probabilities or official policy labels from this mapping.

Validation and limitations

On 464 existing diagnostic inputs, local PyTorch and FP32 ONNX top labels matched; quantization changed seven top labels. This tests export behavior, not held-out accuracy. Browser/WASM behavior is tested separately from desktop ONNX, including the actual Torug worker, rules, thresholds and consent handling.

Diagnostic results show both improvements and regressions against the previous classifier: fewer benign false blocks in several wording groups, but missed harmful requests remain, including sensitivity to uppercase Turkish text. Do not claim that the new model improves every category or guarantees safety.

This is an input signal, not a generative model or an answer-safety score. Torug applies additional rules and thresholds; warnings do not stop generation. Long inputs can lose content through truncation or application segment limits. The model does not evaluate conversation history or provide crisis-support routing. Prompts are processed locally; no hosted inference service is required.

Training data, training code, training/validation/test splits and a permissive model reuse license were not supplied. Diagnostic overlap with training is unknown. No contamination-free result, official TurkBench score or permissive reuse license is claimed. Check applicable rights before redistribution or reuse.