Updated READNE

README.md

@@ -1,3 +1,332 @@
 ---
+language:
+- ru
+- uk
+- en
+tags:
+- text-classification
+- spam-detection
+- ad-filter
+- telegram
+- moderation
+- anti-spam
+- obfuscation
+library_name: transformers
+pipeline_tag: text-classification
+base_model: cointegrated/rubert-tiny2
 license: apache-2.0
 ---
+
+# floxoris/adrash-v0
+
+**Adrash v0** is a compact binary text classification model for detecting advertisements, promo spam, referral spam, Telegram channel promotion, suspicious job spam, and obfuscated ad-like messages.
+
+The model is designed for lightweight moderation systems, especially:
+
+- Telegram bots
+- Telegram groups
+- Telegram Mini Apps
+- marketplaces
+- comment sections
+- chat systems
+- small moderation APIs
+
+**Adrash** means **Ad + Trash**: a small filter that catches advertising garbage before it reaches users.
+
+## Labels
+
+| ID | Label | Meaning |
+|---:|---|---|
+| 0 | `clean` | Normal message |
+| 1 | `ad_spam` | Advertisement, promo, referral spam, job spam, channel promotion, suspicious commercial message |
+
+## What Adrash v0 detects
+
+Adrash v0 is trained to detect messages like:
+
+- Telegram channel promotion
+- referral spam
+- promo-code spam
+- suspicious job offers
+- “work online” spam
+- salary bait messages
+- “write me in DM” spam
+- obfuscated Telegram spam
+- emoji-heavy salary fragments
+- messages with mixed Cyrillic, Latin, and Greek letters
+- messages with hidden Unicode / zero-width characters
+
+Examples of target spam:
+
+```text
+РАБОТА ОНЛАЙН 💰
+Ищу людей в команду на обучение
+Опыт не требуется, всему научу
+ЗП 2000-5000р/день
+Связь: @username
+Подпишись на канал и получи бонус
+```
+
+Obfuscated examples:
+
+```text
+Ηa ceгοдня–зaвтpa нужны 2 чeлοвeκa
+⚠️ЗП в m еcяц 2000💵+
+➡️Uщy людeй в koмaнду на 0бучenиe
+Εcли гοтοвы выйти — пишитe «+» в личныe cοοбщeния
+```
+
+## What Adrash v0 is not for
+
+Adrash v0 is **not** a general safety model.
+
+It is not designed to reliably detect:
+
+- toxicity
+- hate speech
+- violent threats
+- illegal activity
+- self-harm
+- sexual content
+- malware
+- political manipulation
+- general abuse
+
+For those categories, use a separate safety classifier.
+
+## Recommended thresholds
+
+The model outputs probabilities for `clean` and `ad_spam`.
+
+Recommended moderation policy:
+
+| ad_spam score | Action |
+|---:|---|
+| `>= 0.85` | Block / delete |
+| `0.65 - 0.85` | Send to manual moderation |
+| `< 0.65` | Allow |
+
+For production systems, it is better to reduce false positives. Accidentally deleting normal messages is usually worse than missing a small amount of spam.
+
+## Usage with Transformers
+
+```python
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import torch
+
+model_id = "floxoris/adrash-v0"
+
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForSequenceClassification.from_pretrained(model_id)
+
+model.eval()
+
+text = "РАБОТА ОНЛАЙН 💰 ЗП каждый день, пишите в личку"
+
+inputs = tokenizer(
+    text,
+    return_tensors="pt",
+    truncation=True,
+    max_length=160,
+)
+
+with torch.inference_mode():
+    logits = model(**inputs).logits[0]
+    probs = torch.softmax(logits, dim=-1)
+
+clean_score = float(probs[0])
+ad_spam_score = float(probs[1])
+
+label = "ad_spam" if ad_spam_score >= clean_score else "clean"
+
+print({
+    "label": label,
+    "clean": clean_score,
+    "ad_spam": ad_spam_score,
+})
+```
+
+## Usage with pipeline
+
+```python
+from transformers import pipeline
+
+classifier = pipeline(
+    "text-classification",
+    model="floxoris/adrash-v0",
+    tokenizer="floxoris/adrash-v0",
+    return_all_scores=True,
+)
+
+text = "Подпишись на канал и получи бонус"
+result = classifier(text)
+
+print(result)
+```
+
+## Telegram bot moderation example
+
+```python
+def moderation_decision(ad_spam_score: float) -> str:
+    if ad_spam_score >= 0.85:
+        return "block"
+    if ad_spam_score >= 0.65:
+        return "moderate"
+    return "allow"
+```
+
+For Telegram groups, it is recommended to classify a short message buffer from the same user instead of only one isolated message.
+
+Example:
+
+```text
+User sends 3 messages within 20 seconds:
+
+1. РАБОТА ОНЛАЙН 💰
+2. Опыт не требуется, всему научу
+3. Связь: @username
+```
+
+Classifying the combined block is usually more reliable than classifying each fragment separately.
+
+## Training data
+
+Adrash v0 was trained on a mixture of public spam/ham datasets, Telegram-like datasets, synthetic Telegram-style advertisement examples, clean hard-negative examples, and obfuscation-heavy spam samples.
+
+Training sources include:
+
+```text
+thehamkercat/telegram-spam-ham
+mshenoda/spam-messages
+Deysi/spam-detection-dataset
+SetFit/enron_spam
+KSE-RESEARCH-Group/UAReviews
+zefang-liu/phishing-email-dataset
+ucirvine/sms_spam
+SmsSpamCollection
+ScoutieAutoML/russian-news-telegram-dataset
+ScoutieAutoML/cybersecurity_news_telegram_dataset
+```
+
+The training set also includes hard-negative examples such as:
+
+```text
+як зробити реферальну систему в боті?
+потрібно додати кнопку підписатися
+мій Telegram-бот не бачить канал
+скільки коштує реклама в телеграмі?
+це реклама чи нормальний пост?
+```
+
+These examples help reduce false positives on developer, moderation, marketplace, and Telegram-bot related conversations.
+
+## Obfuscation robustness
+
+Adrash v0 was trained with examples containing:
+
+- zero-width Unicode characters
+- Cyrillic / Latin / Greek homoglyph mixing
+- digits used as letters
+- emoji salary fragments
+- short Telegram spam fragments
+- suspicious job-spam patterns
+- mixed-language spam
+- Telegram invite links
+- username/contact bait
+
+Examples:
+
+```text
+⁠⁠⁠⁠⁠⁠⁠⁠⁠РАБОТА О НЛАЙН 💰
+➡️Uщy людeй в koмaнду на 0бучenиe
+⚠️ЗП в m еcяц 2000💵+
+👀 Bсе что нужно - teлeфoн и жeлаnue paб0taть
+✉️ Св⁠язь: @username⁠͏‍
+```
+
+## Evaluation
+
+Replace this section with real metrics from the final training run.
+
+```json
+{
+  "validation": {
+    "eval_precision_ad": "TODO",
+    "eval_recall_ad": "TODO",
+    "eval_f1_ad": "TODO",
+    "eval_false_positive_rate": "TODO",
+    "eval_false_negative_rate": "TODO"
+  },
+  "benchmark": {
+    "benchmark_precision_ad": "TODO",
+    "benchmark_recall_ad": "TODO",
+    "benchmark_f1_ad": "TODO",
+    "benchmark_false_positive_rate": "TODO",
+    "benchmark_false_negative_rate": "TODO"
+  },
+  "hard_test": {
+    "hard_test_precision_ad": "TODO",
+    "hard_test_recall_ad": "TODO",
+    "hard_test_f1_ad": "TODO",
+    "hard_test_false_positive_rate": "TODO",
+    "hard_test_false_negative_rate": "TODO"
+  }
+}
+```
+
+## Limitations
+
+Adrash v0 may still fail on:
+
+- very short fragments without context
+- new spam formats not present in training data
+- messages that require external context
+- mixed moderation categories, such as toxic spam or illegal offers
+- intentionally adversarial text designed to bypass classifiers
+- messages where spam intent is only clear across multiple user messages
+
+For best results, use Adrash v0 together with:
+
+- short user message buffering
+- repeated-message detection
+- link/domain checks
+- rate limits
+- admin review for medium-confidence cases
+
+## Model details
+
+| Field | Value |
+|---|---|
+| Model name | `floxoris/adrash-v0` |
+| Task | Binary text classification |
+| Labels | `clean`, `ad_spam` |
+| Base model | `cointegrated/rubert-tiny2` |
+| Main languages | Russian, Ukrainian, English |
+| Max length used in training | 160 tokens |
+| Framework | Transformers / PyTorch |
+
+## Example output
+
+```json
+{
+  "label": "ad_spam",
+  "clean": 0.0214,
+  "ad_spam": 0.9786
+}
+```
+
+## Citation
+
+```bibtex
+@misc{floxoris_adrash_v0,
+  title={Adrash v0: Compact Advertisement and Spam Filter},
+  author={Floxoris},
+  year={2026},
+  publisher={Hugging Face},
+  howpublished={https://huggingface.co/floxoris/adrash-v0}
+}
+```
+
+## Disclaimer
+
+Adrash v0 is an experimental moderation model. It should not be used as the only moderation layer in high-risk systems. Always test it on your own real messages before production deployment.