README.md

# Floxoris Harmony v1

**Floxoris Harmony v1** is an improved lightweight moderation model for **Russian** and **Ukrainian** text. It continues the idea behind the earlier `Floxoris Harmony v0` release, but presents it as a more mature and refined iteration designed for practical production use where **speed, compact size, and stable binary moderation** matter.

The model is intended for scenarios such as Telegram bots, AI assistants, chat filters, user-generated content pipelines, and message pre-moderation systems. Its goal is simple: provide a fast first-pass toxicity check with minimal deployment overhead.

## What Is New In v1

Compared to the earlier `Floxoris Harmony v0` positioning, **Floxoris Harmony v1** is presented as a cleaner and more production-ready version of the same moderation concept:

- improved and more polished model release
- better suited for practical integration into moderation pipelines
- compact architecture for fast inference
- focused binary safety classification for real-time systems
- clearer output labels: `safe` and `toxic`

Rather than expanding into a large multi-head moderation system, this version keeps the original lightweight philosophy while aiming for a more reliable and streamlined production experience.

## Features

- Binary toxic content classification
- Supports **Russian** and **Ukrainian**
- Small model footprint and fast inference
- Suitable for real-time moderation
- Easy to integrate into lightweight services
- Good fit for bots, assistants, and chat filtering

## Model Details

- **Task:** Binary text classification
- **Architecture:** `BertForSequenceClassification`
- **Model family:** BERT
- **Languages:** Russian, Ukrainian
- **Classes:** `safe`, `toxic`
- **Max sequence length:** 512 tokens
- **Hidden size:** 128
- **Layers:** 2
- **Attention heads:** 2

## Labels

The model returns one of two classes:

- `0` = `safe`
- `1` = `toxic`

## Intended Use

**Floxoris Harmony v1** is intended as a lightweight moderation layer for:

- Telegram bots
- AI assistants
- chat and community platforms
- message pre-filtering
- user input validation before LLM processing
- low-cost production moderation pipelines

Typical usage patterns include:

- filtering incoming messages before sending them to a model or operator
- flagging potentially toxic content for review
- reducing moderation load in high-volume environments
- adding a first-pass safety layer to conversational systems

## Example Usage

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

model_name = "floxoris/floxoris-harmony-v1"

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

id2label = {
    0: "safe",
    1: "toxic",
}

texts = [
    "привіт, як справи?",
    "ти ідіот?",
]

inputs = tokenizer(texts, padding=True, truncation=True, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1)
    preds = torch.argmax(probs, dim=-1)

for text, pred, prob in zip(texts, preds, probs):
    label = id2label[pred.item()]
    confidence = prob[pred.item()].item()
    print(f"{text} -> {label} ({confidence:.4f})")
```

## Example Output

```text
"привіт, як справи?"
-> safe

"ти ідіот?"
-> toxic
```

These examples are illustrative only and should not be treated as a benchmark.

## Limitations

- This is a **binary** moderation model and does not classify toxicity categories
- It may miss subtle harassment, sarcasm, or context-dependent abuse
- It may produce false positives on slang, irony, or emotionally charged language
- Performance can degrade on mixed-language input, heavy misspellings, or niche jargon
- It should be used as a lightweight moderation layer, not as a complete safety system
- Human review is still recommended for high-stakes moderation decisions

## Why This Model

**Floxoris Harmony v1** is built for teams that need:

- simple deployment
- quick inference
- low infrastructure cost
- moderation for Russian and Ukrainian text
- a practical first-stage safety classifier

If you need a compact moderation component instead of a large and expensive multilingual safety stack, this model is designed for that role.

## Summary

**Floxoris Harmony v1** is a new and improved generation of the lightweight moderation approach introduced by `Zire Guard v0`. It keeps the same practical focus on compact binary toxicity detection for Russian and Ukrainian, while presenting a cleaner, more refined, and more production-oriented release for modern moderation pipelines.