Model Card: Mistral-Nemo Debias-Tuned for Geographic Bias in Hate Speech Detection
Model Details
Model Description
This model is a parameter-efficient fine-tuned (PEFT) version of unsloth/Mistral-Nemo-Instruct-2407-bnb-4bit, debias-tuned to mitigate geographic bias in hate speech detection.
The model was trained using a debias tuning strategy, which penalizes inconsistent hate speech classifications when country-specific persona context is added or removed. The objective is to reduce sensitivity to demographic personalization (e.g., “You are from United Kingdom”) while maintaining classification performance.
The model is introduced in:
Piot, P., Martín-Rodilla, P., & Parapar, J. (2025).
“Personalisation or Prejudice? Addressing Geographic Bias in Hate Speech Detection using Debias Tuning in Large Language Models.”
- Developed by: Paloma Piot
- Supervised by: Patricia Martín-Rodilla, Javier Parapar
- Affiliation: IRLab, CITIC Research Centre, Universidade da Coruña
- Model type: Instruction-tuned LLM with PEFT adapter (LoRA-based fine-tuning)
- Language(s): English (evaluation focused on English hate speech detection)
- License: Apache-2.0 (inherits base model constraints where applicable)
- Finetuned from model: unsloth/Mistral-Nemo-Instruct-2407-bnb-4bit
Model Sources
Intended Use
Direct Use
This model is intended for:
- Research on geographic bias in LLM-based hate speech detection
- Studying personalization effects in memory-enabled LLMs
- Evaluating consistency of hate speech classification under persona prompts
- Bias-aware conversational simulations
The model is not designed as a production hate speech moderation system.
Downstream Use
The adapter can be merged into the base model for:
- Fairness-aware classification systems
- Bias evaluation frameworks
- Controlled experiments involving persona-conditioned prompting
Out-of-Scope Use
This model is not suitable for:
- Automated content moderation in high-stakes environments
- Legal or law enforcement decisions
- Deployment without fairness auditing
- Tasks unrelated to hate speech classification
The debiasing method was validated only for hate speech detection and may not generalize to other tasks.
Bias, Risks, and Limitations
Geographic Bias
LLMs may alter their classification outputs depending on country-specific persona prompts (e.g., "You are from United States" vs. "You are from Saudi Arabia").
This model reduces such inconsistencies by penalizing prediction divergence across contexts.
Known Limitations
- Evaluated on a limited set of countries
- Trained for only one epoch to reduce overfitting
- Tested only on hate speech detection
- May reduce sensitivity to legitimate context-specific signals
Societal Risks
If misused, debiasing methods may:
- Mask bias rather than eliminate it
- Produce false confidence in fairness
- Suppress culturally relevant contextual nuance
Ongoing auditing and transparency are recommended.
How to Use
1from transformers import AutoModelForCausalLM, AutoTokenizer
2from peft import PeftModel
3
4base_model_id = "unsloth/Mistral-Nemo-Instruct-2407-bnb-4bit"
5adapter_id = "irlab-udc/Mistral-Nemo-Instruct-2407-Geographic-Debias-Tuning"
6
7tokenizer = AutoTokenizer.from_pretrained(base_model_id)
8base_model = AutoModelForCausalLM.from_pretrained(base_model_id)
9model = PeftModel.from_pretrained(base_model, adapter_id)
10
11prompt = """You are from United Kingdom.
12Is the following text hate speech?
13
14"If a song is sung by a girl there is a 94% chance I'm going to hate it."
15"""
16
17inputs = tokenizer(prompt, return_tensors="pt")
18outputs = model.generate(**inputs, max_new_tokens=50)
19print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Training Details
Training Data
The model was fine-tuned using a subset of the
MetaHate dataset, containing:
- 1.2M+ English hate speech posts aggregated from 36 datasets
- 24,132 instances with country metadata
- Country balancing via augmentation
Training Procedure
A custom consistency-based loss function was used.
For each instance:
- The model classified hate speech with country persona
- The model classified hate speech without persona
- A penalty was applied when predictions were inconsistent
This encourages prediction stability across personalization contexts.
Training Hyperparameters
- Training regime: 4-bit quantized base model + PEFT (LoRA)
- Epochs: 1
- Objective: Classification + consistency penalty
- Precision: 4-bit (bnb quantization via unsloth)
Compute
- Hardware: RTX A6000
- Total compute: 1600 hours (full paper)
- Carbon emissions: 207.36 kgCO₂eq (full paper)
- Carbon efficiency: 0.432 kgCO₂eq/kWh (full paper)
Evaluation
Metrics
- F1-score
- Error analysis (false positives, false negatives, doubtful labels)
- Consistency across persona contexts
Results Summary
The debias-tuned model:
- Reduced prediction inconsistencies between persona and non-persona prompts
- Improved F1 scores in country-specific settings
- Improved robustness across contextual variations
Manual error analysis revealed challenges in:
- Implicit hate
- Idiosyncratic language
- Rare or slang expressions
Technical Specifications
Architecture
- Base: Mistral-Nemo-Instruct-2407 (4-bit quantized)
- Fine-tuning: PEFT (LoRA adapters)
- Debias mechanism: Custom consistency-penalized loss
Objective
Minimize:
- Standard classification loss
- Consistency penalty between persona-conditioned and non-conditioned predictions
Citation
BibTeX
1@misc{piot2025personalisationprejudiceaddressinggeographic,
2 title={Personalisation or Prejudice? Addressing Geographic Bias in Hate Speech Detection using Debias Tuning in Large Language Models},
3 author={Paloma Piot and Patricia Martín-Rodilla and Javier Parapar},
4 year={2025},
5 eprint={2505.02252},
6 archivePrefix={arXiv},
7 primaryClass={cs.CL},
8 url={https://arxiv.org/abs/2505.02252},
9}
Contact
Paloma Piot
IRLab, CITIC Research Centre
Universidade da Coruña
paloma.piot@udc.es 
