ArchMC/_model_1

1---2license: mit3language: en4tags:5  - text-classification6  - toxicity7  - moderation8  - chat9  - bert10  - pytorch11  - onnx12datasets:13  - dormlab/chat-corpus14metrics:15  - accuracy16  - f117  - precision18  - recall19pipeline_tag: text-classification20---21 22# Toxic Chat Moderation23 24Binary classifier for real-time chat moderation. Flags toxic, hateful, harassing,25sexually explicit, and otherwise inappropriate messages in gaming and social chat.26 27Based on  fine-tuned on 300K labeled chat messages.28 29## Quick use30 31 32 33## Performance34 35| Metric | Score |36|--------|-------|37| Accuracy | 0.9768 |38| F1 | 0.9768 |39| Precision | 0.9643 |40| Recall | 0.9897 |41 42ONNX INT8 latency: ~1-3ms on Apple Silicon (CoreML/MPS).43 44## Training45 46- **Architecture**: bert-base-uncased (110M params), 2 labels (clean/toxic)47- **Hardware**: Apple Silicon Mac Mini (MPS), single-node48- **Data**: 153K messages (122,688 train / 15,336 val / 15,336 test)49- **Framework**: PyTorch, HuggingFace Trainer50- **Export**: ONNX dynamic INT8 quantization (105 MB)51 52## Variants53 54This repo provides two model formats:55-  — full PyTorch weights for use with usage: transformers <command> [<args>]56 57positional arguments:58  {chat,convert,download,env,run,serve,add-new-model-like,add-fast-image-processor}59                        transformers command helpers60    convert             CLI tool to run convert model from original author61                        checkpoints to Transformers PyTorch checkpoints.62    run                 Run a pipeline through the CLI63    serve               CLI tool to run inference requests through REST and64                        GraphQL endpoints.65 66options:67  -h, --help            show this help message and exit68-  — ONNX INT8 quantized for fast inference on CPU/CoreML69 70## Label mapping71 72| Label | Meaning |73|--------|---------|74| 0 | Clean — allow |75| 1 | Toxic — block/flag |76