berkeruveyik/toxic-speech-finetune-with-gemma-3-1b-v1

Toxic Speech Structured Analysis (Gemma 3 1B IT)

Model Description

This model is fine-tuned from google/gemma-3-1b-it for structured toxic speech analysis in English and Turkish.

Instead of returning open-ended chat responses, it is trained to extract:

• number_of_bad_words
• tags
• insult_word

The model is designed for short social text such as comments, captions, and forum-like messages.

Task

Given an input text, the model predicts:

• The number of offensive expressions.
• The toxicity label(s).
• The insulting/offensive words or phrases.

Supported label set:

• Strong Insult
• Toxic
• Sarcastic-Mocking
• Strong Insult|Threat
• Mobbing
• Mild Insult
• Discriminatory
• Threat
• Passive-Aggressive
• Strong Insult|Discriminatory

Training Data

• Dataset: berkeruveyik/toxic-speech-annotated-dataset
• Languages: English and Turkish
• Supervision format: conversational (user input + assistant structured target)

Example target format:

number_of_bad_words: 2
tags: Strong Insult
insult_word: stupid, ugly

Training Procedure

The model was trained with supervised fine-tuning using TRL (SFTTrainer) on structured conversation data.

Typical training setup:

• max_length=512
• num_train_epochs=10
• per_device_train_batch_size=8
• per_device_eval_batch_size=8
• gradient_accumulation_steps=2
• learning_rate=2e-5
• lr_scheduler_type=cosine
• warmup_ratio=0.1
• weight_decay=0.01
• save_strategy=epoch
• eval_strategy=epoch
• load_best_model_at_end=True

Inference

For better format consistency, use deterministic decoding (do_sample=False).

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

MODEL_ID = "YOUR_USERNAME/YOUR_MODEL_REPO"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, device_map="auto", dtype="auto")

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)

text = "You are useless and disgusting."
messages = [{"role": "user", "content": text}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

out = pipe(prompt, max_new_tokens=128, do_sample=False)
print(out[0]["generated_text"][len(prompt):])

Intended Use

• Toxicity signal extraction for moderation workflows
• Annotation support
• Safety analysis prototypes for EN/TR text

Limitations

• May miss implicit abuse, sarcasm, or context-heavy harassment.
• Extracted insult words can be incomplete or noisy.
• Performance may drop on long inputs, slang drift, or unseen domains.
• Should not be the sole basis for legal, HR, or punitive decisions.

Responsible Use

• Keep human review for high-stakes cases.
• Monitor false positives and false negatives.
• Re-evaluate regularly across both languages and target domains.

Suggested Evaluation Metrics

• Tag-level Precision / Recall / F1
• Exact-match rate for full structured output
• number_of_bad_words MAE
• Parse-valid output rate

Credits

• Base model: google/gemma-3-1b-it
• Dataset: berkeruveyik/toxic-speech-annotated-dataset
• License: Apache-2.0