README.md

# Floxoris Harmony v1.1

**Floxoris Harmony v1.1** is a lightweight moderation model for fast binary toxicity detection in Russian and Ukrainian text.

This version is a continued fine-tuning update of **Floxoris Harmony v1**, focused on improving detection of **mild toxicity**, short rude phrases, and everyday aggressive messages while keeping the model compact and fast.

The model is intended for scenarios where **low latency, small size, and simple deployment** matter, such as Telegram bots, chat moderation systems, AI assistants, community tools, and first-pass safety filters.

## What Is New In v1.1

Compared to `Floxoris Harmony v1`, this release focuses on better handling of short and mild toxic messages.

Examples of targeted improvements:

- better detection of short rude phrases
- improved sensitivity to mild toxicity
- stronger Russian and Ukrainian moderation behavior
- better handling of direct insults and aggressive commands
- continued support for fast binary classification
- same simple output labels: `safe` and `toxic`

This version was trained as a **behavioral patch**, not as a completely new architecture.

## Model Task

The model performs binary text classification:

| Class | Label |
|---|---|
| `0` | `safe` |
| `1` | `toxic` |

It is designed to answer a simple question:

> Is this message safe or toxic?

## Intended Use

Floxoris Harmony v1.1 is suitable for:

- Telegram bot moderation
- chat message filtering
- AI assistant safety checks
- community moderation tools
- lightweight API moderation
- first-stage toxicity detection
- Russian/Ukrainian text safety classification

It works best as a fast first-pass classifier before more complex moderation logic.

## Example Usage

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "floxoris/harmony-v1.1"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

text = "заткнись"

inputs = tokenizer(
    text,
    return_tensors="pt",
    truncation=True,
    padding=True,
    max_length=128
)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)[0]

safe_score = probs[0].item()
toxic_score = probs[1].item()

label = "toxic" if toxic_score > safe_score else "safe"

print({
    "label": label,
    "safe_score": round(safe_score, 4),
    "toxic_score": round(toxic_score, 4)
})