content-safety-xlm-roberta-large

A multilingual seven-head content-safety classifier fine-tuned from FacebookAI/xlm-roberta-large by HikmaAI.

Model description

• Task: independent multi-label sequence classification
• Version: 1.1.0
• Base model: FacebookAI/xlm-roberta-large
• Languages: 11 (en, vi, hi, th, zh, ja, ru, ar, sv, es, it)
• Heads: hate, abuse, violence, self_harm, sexual_harm, misconduct, dangerous_capability
• Activation: one sigmoid per head; never softmax
• Maximum length: 512 tokens; production preprocessing uses head-tail truncation
• Export formats: ONNX FP32 + FP16 + INT8 (CPU dynamic quantization)

The model intentionally emits no aggregate safe, threat, harmful, or block decision. Policy aggregation and enforcement belong to the consuming application.

Held-out test performance

Metric	Score
macro_average_precision	0.8784
macro_f1	0.8038
worst_head_average_precision	0.8056
worst_head_f1	0.7541
micro_f1	0.8025
exact_match	0.9157

Head	AP	Precision	Recall	F1
hate	0.8835	0.7829	0.8330	0.8072
abuse	0.8955	0.7695	0.8547	0.8098
violence	0.8724	0.7245	0.8509	0.7826
self_harm	0.8056	0.6891	0.8326	0.7541
sexual_harm	0.8578	0.7885	0.8218	0.8048
misconduct	0.8859	0.7239	0.8483	0.7812
dangerous_capability	0.9482	0.9387	0.8407	0.8870

The test labels are teacher-generated and are not independent human gold. Legacy-safe training records were assigned all-zero vectors. Metrics therefore measure alignment to the documented training taxonomy and require independent human validation before high-stakes deployment.

Calibrated thresholds

Head	Threshold
hate	0.853
abuse	0.794
violence	0.841
self_harm	0.968
sexual_harm	0.865
misconduct	0.701
dangerous_capability	0.996

Thresholds are bound to the exported model configuration hash. Apply each threshold only to its corresponding sigmoid score.

Usage (ONNX)

import json
import torch
from huggingface_hub import hf_hub_download
from optimum.onnxruntime import ORTModelForSequenceClassification
from transformers import AutoTokenizer

repo_id = "HikmaAI/content-safety-xlm-roberta-large"
subfolder = "onnx/fp16"
model = ORTModelForSequenceClassification.from_pretrained(repo_id, subfolder=subfolder)
tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder="tokenizer")
thresholds = json.load(open(hf_hub_download(repo_id, "thresholds.json")))
heads = ["hate", "abuse", "violence", "self_harm", "sexual_harm",
         "misconduct", "dangerous_capability"]

inputs = tokenizer("Example text", return_tensors="pt", truncation=True, max_length=512)
scores = torch.sigmoid(model(**inputs).logits)[0].tolist()
result = {
    head: {"score": score, "detected": score >= thresholds["heads"][head]["threshold"]}
    for head, score in zip(heads, scores)
}

For exact parity with training on texts longer than 512 tokens, retain the first 256 and last 256 token IDs before padding.

CPU INT8 export

The CPU artifact is onnx/int8/model_quantized.onnx (562,815,147 bytes, 562.8 decimal MB), with tokenizer.json, tokenizer metadata, config.json and thresholds.json in the same folder. All weights are inline; no external-data sidecar is needed for INT8. FP16 and FP32 keep their existing external-data files.

The export uses ONNX Runtime 1.29.0 dynamic quantization, per-channel QInt8 weights, from FP32 source revision bebab7517b618b820a39af574f2919a151bc6fa0. Use ONNX Runtime 1.29.0 for the validated inference behavior. Compute one sigmoid per category and preserve the 512-token head/tail preprocessing described above. See onnx/int8/preparation.json for hashes and provenance.

Local checks on macOS arm64 covered 42 hand-authored probes: benign examples in all 11 languages, harmful requests across the seven categories in English, Italian and Arabic, Unicode, and long-text preprocessing. Token IDs matched the published slow tokenizer and the native Go tokenizer exactly. No category threshold decision changed between FP32 and INT8 on those probes; Python and the native Go runtime produced identical INT8 scores with ONNX Runtime 1.29.0.

These are smoke checks, not a labelled quality benchmark. FP32/INT8 probability differences reached approximately 0.128 on this corpus. The thresholds are copied reference values and have not been recalibrated for INT8; the held-out metrics above were not remeasured for this export. Full labelled quality evaluation and real target amd64 qualification remain pending before production enforcement.

Training

• Fine-tuning: LoRA rank 16 on attention query/value, merged for export
• Epochs: 3
• Learning rate: 0.0002
• Effective batch size: 16

License

Apache-2.0

Citation

@misc{hikmaai-content_safety-2026,
  title={content-safety-xlm-roberta-large},
  author={HikmaAI},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/HikmaAI/content-safety-xlm-roberta-large}
}