Intended use

The model receives:

• A user complaint or description of a social problem
• The requested response language
• A response style
• A topic label

It generates a concise response that:

• Acknowledges the underlying issue
• Converts frustration into concrete next steps
• Follows the requested tone
• Responds in English or Ukrainian
• Avoids assuming the user's emotional state
• Attempts to avoid unsupported facts, organizations, locations, and links

The supported response styles in the training dataset are:

• normal
• polite
• funny
• sarcastic
• rude

Role in Hate-2-Action

The complete Hate-2-Action pipeline:

1. Receives a complaint or problem from a Telegram user.
2. Detects the language and conversation intent.
3. Extracts problems and possible solution concepts.
4. Uses embeddings and vector similarity to find relevant NGOs and projects.
5. Passes the resulting context to the final response generator.
6. Produces a short, styled, actionable answer.

This model is intended for step 6: final natural-language answer generation.

Training

The model was fine-tuned from:

unsloth/gemma-4-e2b-it-unsloth-bnb-4bit

Fine-tuning was performed using:

• Unsloth
• Hugging Face TRL

Training data

The model was trained on dataset from the Hate-2-Action project.

The dataset contains 2,000 supervised examples:

• 1,000 English examples
• 1,000 Ukrainian examples
• 400 examples for each response style
• 200 examples for each of 10 topic categories

The dataset builder also produced nominal splits:

• Train: 1,800 examples
• Validation: 100 examples
• Test: 100 examples

Recommended deployment

For production use, provide the model with verified organization and project context retrieved by the Hate-2-Action database.

Generated links, factual claims, and organization details should be validated before being shown to users.

This model is an experimental project-specific checkpoint and should be evaluated against a new, separately created test set before replacing the existing production LLM.