sage

Local multi-label content-moderation model for chat and roleplay.

• 137M parameters (fine-tune of jinaai/jina-embeddings-v2-base-en)
• 7 categories: nsfw, violence, harassment, hate_speech, self_harm, grooming, sexual_minors
• Accepts a single message or up to 16 turns of conversation; classifies the last turn
• INT8 ONNX artifact ~35 MB, CPU inference ~30–50 ms at 1024 tokens
• Python and Rust inference libraries; no network dependency

Status — pre-1.0. Training, evaluation, and first model release are in progress.

pip install sage-moderation
# with training/export extras:
pip install "sage-moderation[train,export]"

Rust crate:

[dependencies]
sage = { git = "https://github.com/MegalithOfficial/sage" }

from sage.inference import Sage

model = Sage.from_onnx("sage-v1-int8.onnx")

model.check("some text")

model.check([
    {"role": "user", "text": "earlier message"},
    {"role": "char", "text": "reply"},
    {"role": "user", "text": "message to classify"},
])

Rust:

use sage::{Sage, Turn};

let model = Sage::from_onnx("sage-v1-int8.onnx", "tokenizer.json")?;
let result = model.check(&[
    Turn::user("earlier message"),
    Turn::char("reply"),
    Turn::user("message to classify"),
])?;

Output shape:

{
  "flagged": true,
  "categories": {
    "nsfw":          {"score": 0.03, "flagged": false},
    "violence":      {"score": 0.08, "flagged": false},
    "harassment":    {"score": 0.06, "flagged": false},
    "hate_speech":   {"score": 0.04, "flagged": false},
    "self_harm":     {"score": 0.02, "flagged": false},
    "grooming":      {"score": 0.71, "flagged": true},
    "sexual_minors": {"score": 0.04, "flagged": false}
  }
}

sage/          python package (schema, tokenizer, model, inference, loss)
training/      data pipeline, trainer, onnx exporter
rust/sage-rs/  rust inference crate
benchmarks/    adversarial test suites
notebooks/     colab notebooks for training and synthesis
scripts/       command-line review + eval tools
docs/          architecture and pipeline docs
tests/         pytest

python -m venv venv
source venv/bin/activate
pip install -e '.[train,export,dev]'
make check

python -m training.data.aggregate --out data/processed --sources all --limit-per-source 50000
for s in train val test; do
  python -m training.data.trajectorize_cli --in data/processed/$s.jsonl --out data/processed/$s.traj.jsonl
done
python -m training.train \
  --train data/processed/train.traj.jsonl \
  --val   data/processed/val.traj.jsonl \
  --out   checkpoints/sage-v1
python -m training.export_onnx \
  --checkpoint checkpoints/sage-v1/best.pt \
  --out-fp32 artifacts/sage-v1-fp32.onnx \
  --out-int8 artifacts/sage-v1-int8.onnx

See docs/training.md, docs/data_pipeline.md, docs/architecture.md.

Seven openly-licensed, non-gated sources. See DATA_LICENSES.md.

Rare categories are augmented with reviewer-approved LLM-generated conversations. Raw synthetic data is not distributed.

• Code: Apache 2.0 (LICENSE)
• Model weights: subject to constituent dataset licenses