# Model Card for Model ID Lynx is an open-source hallucination evaluation model. Patronus-Lynx-8B-Instruct-v1.1 was trained on a mix of datasets including CovidQA, PubmedQA, DROP, RAGTruth. The datasets contain a mix of hand-annotated and synthetic data. The maximum sequence length is 128000 tokens. ## Model Details - **Model Type:** Patronus-Lynx-8B-Instruct-v1.1 is a fine-tuned version of meta-llama/Meta-Llama-3.1-8B-Instruct model. - **Language:** Primarily English - **Developed by:** Patronus AI - **Paper:** [https://arxiv.org/abs/2407.08488](https://arxiv.org/abs/2407.08488) - **License:** [https://creativecommons.org/licenses/by-nc/4.0/](https://creativecommons.org/licenses/by-nc/4.0/) ### Model Sources - **Repository:** [https://github.com/patronus-ai/Lynx-hallucination-detection](https://github.com/patronus-ai/Lynx-hallucination-detection) ## How to Get Started with the Model Lynx is trained to detect hallucinations in RAG settings. Provided a document, question and answer, the model can evaluate whether the answer is faithful to the document. To use the model, we recommend using the following prompt: ``` PROMPT = """ Given the following QUESTION, DOCUMENT and ANSWER you must analyze the provided answer and determine whether it is faithful to the contents of the DOCUMENT. The ANSWER must not offer new information beyond the context provided in the DOCUMENT. The ANSWER also must not contradict information provided in the DOCUMENT. Output your final verdict by strictly following this format: "PASS" if the answer is faithful to the DOCUMENT and "FAIL" if the answer is not faithful to the DOCUMENT. Show your reasoning. -- QUESTION (THIS DOES NOT COUNT AS BACKGROUND INFORMATION): {question} -- DOCUMENT: {context} -- ANSWER: {answer} -- Your output should be in JSON FORMAT with the keys "REASONING" and "SCORE": {{"REASONING": , "SCORE": }} """ ``` The model will output the score as 'PASS' if the answer is faithful to the document or FAIL if the answer is not faithful to the document. ## Inference To run inference, you can use HF pipeline: ``` model_name = 'PatronusAI/Llama-3-Patronus-Lynx-8B-Instruct-v1.1' pipe = pipeline( "text-generation", model=model_name, max_new_tokens=600, device="cuda", return_full_text=False ) messages = [ {"role": "user", "content": prompt}, ] result = pipe(messages) print(result[0]['generated_text']) ``` Since the model is trained in chat format, ensure that you pass the prompt as a user message. For more information on training details, refer to our [ArXiv paper](https://arxiv.org/abs/2407.08488). ## Evaluation The model was evaluated on [PatronusAI/HaluBench](https://huggingface.co/datasets/PatronusAI/HaluBench). | Model | HaluEval | RAGTruth | FinanceBench | DROP | CovidQA | PubmedQA | Overall | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | GPT-4o | 87.9% | 84.3% | 85.3% | 84.3% | 95.0% | 82.1% | 86.5% | | GPT-4-Turbo | 86.0% | 85.0% | 82.2% | 84.8% | 90.6% | 83.5% | 85.0% | | GPT-3.5-Turbo | 62.2% | 50.7% | 60.9% | 57.2% | 56.7% | 62.8% | 58.7% | | Claude-3.5-Sonnet | 84.5% | 79.1% | 69.3% | 69.7% | 70.8% |84.8% |83.7%| | RAGAS Faithfulness | 70.6% | 75.8% | 59.5% | 59.6% | 75.0% | 67.7% | 66.9% | | Mistral-Instruct-7B | 78.3% | 77.7% | 56.3% | 56.3% | 71.7% | 77.9% | 69.4% | | Llama-3-Instruct-8B | 83.1% | 80.0% | 55.0% | 58.2% | 75.2% | 70.7% | 70.4% | | Llama-3-Instruct-70B | 87.0% | **83.8%** | 72.7% | 69.4% | 85.0% | 82.6% | 80.1% | | Lynx (8B) | 85.7% | 80.0% | 72.5% | **77.8%** | 96.3% | 85.2% | 82.9% | | Lynx v1.1 (8B) | **87.3%** | 79.9% | **75.6%** | 77.5% | **96.9%** | **88.9%** | **84.3%** | ## Citation If you are using the model, cite using ``` @article{ravi2024lynx, title={Lynx: An Open Source Hallucination Evaluation Model}, author={Ravi, Selvan Sunitha and Mielczarek, Bartosz and Kannappan, Anand and Kiela, Douwe and Qian, Rebecca}, journal={arXiv preprint arXiv:2407.08488}, year={2024} } ``` ## Model Card Contact [@sunitha-ravi](https://huggingface.co/sunitha-ravi) [@RebeccaQian1](https://huggingface.co/RebeccaQian1) [@presidev](https://huggingface.co/presidev)