🧠 Toxicity_model_Llama_3.1_8B – Spanish Toxicity Classifier (Instruction-Tuned)

📌 Model Description

This model is an instruction-tuned version of LLaMA 3.1 8B Instruct, specifically trained to classify the toxicity level of Spanish-language user comments on news articles. It distinguishes between three categories:

• Non-toxic
• Slightly toxic
• Toxic

The model follows instruction-based prompts and returns a single classification label in response.

---

📂 Training Data

The model was fine-tuned on the SocialTOX dataset, a collection of Spanish-language comments annotated for varying levels of toxicity. These comments come from news platforms and represent real-world scenarios of online discourse.

---

💬 Instruction Format

The model expects input in a structured conversational format like the one below:

[
  {
    "role": "system",
    "content": "You are an expert in detecting toxicity in comments, and your goal is to classify comments based on their level of toxicity. The comments were made on news articles. The toxicity categories are:
    Slightly toxic: Comments that express sarcasm, irony, or rhetorical questions, but do not directly attack or degrade others.
    Toxic: Comments that contain derogatory or pejorative language, inappropriate jokes, fearmongering, denial of facts, threats, personal attacks, insults, degradation, or racist or sexist language. Only classify a comment as “toxic” if it contains clear attack language, direct insults, or demeaning references.
    Non-toxic: Neutral or critical comments that do not include Toxic or Slightly toxic elements. Note that negative or critical comments (those with a serious or discontented tone) are Not toxic or Slightly toxic unless they meet the criteria of the categories above.
    Please write the corresponding category immediately after the word 'answer.' In case of doubt between two labels, choose the one with the lowest or no toxicity level."
  },
  {
    "role": "user",
    "content": "Text: "Narco-Bolivarian Communism"
  },
  {
    "role": "assistant",
    "content": "Toxic"
  }
]

Training hyperparameters

• epochs: 3
• learning_rate: 1e-5
• beta1: 0.9
• beta2: 0.95
• weight_decay: 0.1
• batch_size global: 4
• micro_batch_size: 1
• lr_warmup_steps: 100
• max_seq_length: 512

📊 Evaluation

The model was evaluated on a held-out test set of 968 manually annotated comments. Below are the confusion matrix and classification metrics:

🧮 Confusion Matrix

	Non-toxic	Slightly Toxic	Toxic
Non-toxic	325	52	62
Slightly Toxic	116	257	8
Toxic	67	12	69

📈 Classification Report

Class	Precision	Recall	F1-score	Support
Non-toxic	0.8006	0.6745	0.7322	381
Slightly toxic	0.6398	0.7403	0.6864	439
Toxic	0.4964	0.4662	0.4808	148
Accuracy			0.6725	968
Macro average	0.6456	0.6270	0.6331	968
Weighted average	0.6812	0.6725	0.6730	968

Macro F1-score: 0.6331