README.md

---
license: gemma
language:
- en
tags:
- truthfulqa
- llm-judge
- hitz
- gemma
- en
- truth-judge
datasets:
- HiTZ/truthful_judge
base_model: google/gemma-2-9b-it
---

# Model Card for HiTZ/gemma-2-9b-it-en-truth-judge

This model card is for a judge model fine-tuned to evaluate truthfulness, based on the work "Truth Knows No Language: Evaluating Truthfulness Beyond English".

## Model Details

### Model Description

This model is an LLM-as-a-Judge, fine-tuned from `google/gemma-2-9b-it` to assess the truthfulness of text generated by other language models. The evaluation framework and findings are detailed in the paper "Truth Knows No Language: Evaluating Truthfulness Beyond English." The primary goal of this work is to extend truthfulness evaluations beyond English, covering Basque, Catalan, Galician, and Spanish.

- **Developed by:** Blanca Calvo Figueras, Eneko Sagarzazu, Julen Etxaniz, Jeremy Barnes, Pablo Gamallo, Iria De Dios Flores, Rodrigo Agerri.
- **Affiliations:** HiTZ Center - Ixa, University of the Basque Country, UPV/EHU; Elhuyar; Centro de Investigación en Tecnoloxías Intelixentes (CiTIUS), Universidade de Santiago de Compostela; Departament de Traducció i Ciències del Llenguatge, Universitat Pompeu Fabra.
- **Funded by:** MCIN/AEI/10.13039/501100011033 projects: DeepKnowledge (PID2021-127777OB-C21) and by FEDER, EU; Disargue (TED2021-130810B-C21) and European Union NextGenerationEU/PRTR; DeepMinor (CNS2023-144375) and European Union NextGenerationEU/PRTR; NÓS-ILENIA (2022/TL22/0021533). Xunta de Galicia: Centro de investigación de Galicia accreditation 2024-2027 ED431G-2023/04. UPV/EHU PIF22/84 predoc grant (Blanca Calvo Figueras). Basque Government PhD grant PRE_2024_2_0028 (Julen Etxaniz). Juan de la Cierva contract and project JDC2022-049433-I (Iria de Dios Flores), financed by the MCIN/AEI/10.13039/501100011033 and the European Union “NextGenerationEU”/PRTR.
- **Shared by:** HiTZ Center
- **Model type:** LLM-as-a-Judge, based on `Gemma2`
- **Language(s) (NLP):** Fine-tuned to judge outputs in `English`. The underlying TruthfulQA-Multi benchmark, used for context, covers English, Basque, Catalan, Galician, and Spanish.
- **License:** The base model `google/gemma-2-9b-it` is governed by the Gemma license. The fine-tuning code, this model's weights, and the TruthfulQA-Multi dataset are publicly available under Apache 2.0.
- **Finetuned from model:** `google/gemma-2-9b-it`

### Model Sources

- **Repository (for the project and fine-tuning code):** `https://github.com/hitz-zentroa/truthfulqa-multi`
- **Paper:** "Truth Knows No Language: Evaluating Truthfulness Beyond English" (`https://arxiv.org/abs/2502.09387`)
- **Dataset (TruthfulQA-Multi):** `https://huggingface.co/datasets/HiTZ/truthful_judge`

## Uses

### Direct Use

This model is intended for direct use as an LLM-as-a-Judge. It takes a question, a reference answer, and a model-generated answer as input, and outputs a judgment on the truthfulness of the model-generated answer. This is particularly relevant for evaluating models on the TruthfulQA benchmark, specifically for English.

### Downstream Use

This judge model could potentially be used as a component in larger systems for content moderation, automated fact-checking research, or as a basis for further fine-tuning on more specific truthfulness-related tasks or domains.

### Out-of-Scope Use

This model is not designed for:
- Generating general-purpose creative text or dialogue.
- Providing factual information directly (it judges, it doesn't assert).
- Use in safety-critical applications without thorough validation.
- Any application intended to deceive or spread misinformation.
The model's judgments are based on its training and may not be infallible.

## Bias, Risks, and Limitations

The model's performance and biases are influenced by its base model (`google/gemma-2-9b-it`) and the TruthfulQA dataset. Key points from "Truth Knows No Language: Evaluating Truthfulness Beyond English":
- **Language Discrepancies:** LLMs (and thus judges based on them) may perform differently across languages. The paper found English performance to be generally highest.
- **Anglocentricity:** The original TruthfulQA benchmark has English-centric knowledge and cultural contexts. This model is specifically for English.
- **Context and Time Dependency:** Universal knowledge questions are handled better across languages than context- and time-dependent ones.
- **Informativeness:** While this is a "truth-judge", informativeness is a critical component of truthfulness assessment.
- **Imitative Falsehoods:** TruthfulQA focuses on "imitative falsehoods" – statements that are false but sound plausible or mimic common misconceptions.

### Recommendations

- Users should be aware of the potential for varying performance across different types of questions (e.g., universal vs. culturally specific).
- For critical applications, judgments from this model should be cross-verified, ideally with human oversight.
- Consider the limitations regarding cultural and temporal variability when interpreting results.

## How to Get Started with the Model

This model can be loaded using the Hugging Face `transformers` library.
```python
# Example (conceptual, actual usage depends on task setup)
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "HiTZ/gemma-2-9b-it-en-truth-judge"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# For LLM-as-a-Judge, you would typically format input like:
# prompt = f"Question: {question}\\nReference Answer: {ref_answer}\\nModel Answer: {model_answer}\\nIs the model answer truthful? Judge:"
# inputs = tokenizer(prompt, return_tensors="pt")
# outputs = model.generate(**inputs) # Adjust generation parameters as needed
# judgment = tokenizer.decode(outputs[0], skip_special_tokens=True)
```
Refer to the project repository (`https://github.com/hitz-zentroa/truthfulqa-multi`) for specific examples of how judge models were used in the evaluation.

## Training Details

### Training Data

The model was fine-tuned on a dataset derived from the original English TruthfulQA benchmark \cite{lin-etal-2022-truthfulqa}.
- **Dataset Link:** `https://huggingface.co/datasets/HiTZ/truthful_judge` (original English portion)
- **Training Data Specifics:** Trained on English data for truth judging.

### Training Procedure

The model was fine-tuned as an LLM-as-a-Judge. The methodology was adapted from the original TruthfulQA paper \cite{lin-etal-2022-truthfulqa}, where the model learns to predict whether an answer is truthful given a question and reference answers.

#### Preprocessing

Inputs were formatted to present the judge model with a question, correct answer(s), and the answer to be judged, prompting it to assess truthfulness.

#### Training Hyperparameters

- **Training regime:** `bfloat16` mixed precision
- **Base model:** `google/gemma-2-9b-it`
- **Epochs:** 5
- **Learning rate:** 0.01
- **Batch size:** Refer to project code
- **Optimizer:** Refer to project code
- **Transformers Version:** `4.44.2`

## Evaluation

### Testing Data, Factors & Metrics

#### Testing Data

The model's evaluation methodology is described in "Truth Knows No Language: Evaluating Truthfulness Beyond English," using questions from the TruthfulQA-Multi dataset (English portion).

#### Factors

- **Language:** English.
- **Model Type (of models being judged):** Base and instruction-tuned LLMs.
- **Evaluation Metric:** Correlation of LLM-as-a-Judge scores with human judgments on truthfulness; comparison with multiple-choice metrics (MC2).

#### Metrics

- **Primary Metric:** Spearman correlation between the judge model's scores and human-annotated scores for truthfulness.
- The paper found that LLM-as-a-Judge (like this model) correlates more closely with human judgments than multiple-choice metrics. For the general Gemma-2-9b-it judge trained on all languages (MT data), Kappa was 0.74 for English (Table 3 in paper).

### Results

#### Summary

As reported in "Truth Knows No Language: Evaluating Truthfulness Beyond English":
- LLMs generally perform best in English.
- LLM-as-a-Judge models demonstrated a stronger correlation with human judgments compared to MC2 metrics.
- This specific model (`gemma9b_instruct_truth_judge`) is one of the judge models fine-tuned for the experiments. Refer to Table 3 in the paper for Judge-LLM performance (Gemma 2 9B IT was the base for the best Judge-LLM).

## Technical Specifications

### Model Architecture and Objective

The model is based on the `Gemma2` architecture (`Gemma2ForCausalLM`). It is a Causal Language Model fine-tuned with the objective of acting as a "judge" to predict the truthfulness of answers to questions, particularly those designed to elicit imitative falsehoods.
- **Hidden Size:** 3584
- **Intermediate Size:** 14336
- **Num Attention Heads:** 16
- **Num Hidden Layers:** 42
- **Num Key Value Heads:** 8
- **Vocab Size:** 256000

### Compute Infrastructure

- **Hardware:** Refer to project for details.
- **Software:** PyTorch, Transformers `4.44.2`

## Citation

**Paper:**
```bibtex
@inproceedings{calvo-etal-2025-truthknowsnolanguage,
    title = "Truth Knows No Language: Evaluating Truthfulness Beyond English",
    author = "Calvo Figueras, Blanca and Sagarzazu, Eneko and Etxaniz, Julen and Barnes, Jeremy and Gamallo, Pablo and De Dios Flores, Iria and Agerri, Rodrigo",
    year={2025},
    eprint={2502.09387},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
    url={https://arxiv.org/abs/2502.09387}
}
```

## More Information

For more details on the methodology, dataset, and findings, please refer to the full paper "Truth Knows No Language: Evaluating Truthfulness Beyond English" and the project repository: `https://github.com/hitz-zentroa/truthfulqa-multi`.

## Model Card Authors

This model card was generated based on information from the paper "Truth Knows No Language: Evaluating Truthfulness Beyond English" by Blanca Calvo Figueras et al., and adapted from the Hugging Face model card template. Content populated by GitHub Copilot.

## Model Card Contact

For questions about the model or the research, please contact:
- Blanca Calvo Figueras: `[email protected]`
- Rodrigo Agerri: `[email protected]`