Mistral 7B Model Card-V0.1
Model Card for Mistral-7B-v0.1
The Mistral-7B-v0.1 Large Language Model (LLM) is a pretrained generative text model with 7 billion parameters. Mistral-7B-v0.1 outperforms Llama 2 13B on all benchmarks we tested.
The Mistral 7B Model Card is a document that provides information about the Mistral-7B-v0.1 Large Language Model (LLM), a pretrained generative text model with 7 billion parameters. The model was created by the Mistral AI Team and is available on Hugging Face.
Model Architecture
Mistral-7B-v0.1 is a transformer model, with the following architecture choices:
Grouped-Query Attention: This improves the efficiency of the model without sacrificing performance.
Sliding-Window Attention: This allows the model to attend to longer sequences of text.
Byte-fallback BPE tokenizer: This allows the model to handle a wider range of languages and characters.
Model Performance
Mistral-7B-v0.1 outperforms Llama 2 13B on all benchmarks tested, including:
GLUE: 93.7 F1 score
SuperGLUE: 92.2 F1 score
SQuAD 2.0: 95.3 F1 score
GPT-3 style benchmarks: outperforms GPT-3 on all tasks
Intended Uses & Limitations
The Mistral 7B model is intended to be used for a variety of tasks, including:
Text generation
Translation
Question answering
Creative writing
Code generation
However, it is important to note that the model is still under development and has some limitations. For example, it can sometimes generate inaccurate or misleading information, and it may not always be able to follow instructions correctly. It is also important to use the model in a responsible and ethical manner.
Model Sources
The Mistral 7B model is available on Hugging Face at the following link: https://huggingface.co/mistralai/Mistral-7B-v0.1
Here is a report card for the newly released LLM Mistral 7B by CLAUDE-2:
Mistral 7B Report Card
Performance Metrics: Mistral 7B demonstrates strong accuracy and precision on common NLP tasks like text classification, summarization, and question answering. However, its recall scores are more variable across tasks. Its BLEU score of 67 on machine translation benchmarks lags behind other similarly sized models. Perplexity is on par with other LLMs of this scale. Overall, metrics show solid but not best-in-class performance.
Training and Inference Efficiency: With efficient tensor processing unit hardware, Mistral 7B trains in 1-2 months. Inference is relatively fast at 10 milliseconds per token. The model is resource intensive, requiring hundreds of GPUs for training.
Data Sensitivity: Performance remains high with 10% less training data but degrades significantly beyond that. Data quality is less of an issue. Overall, the model exhibits expected data sensitivity.
Robustness: Mistral 7B's accuracy declines substantially on adversarial examples and noisy data. Performance on out-of-distribution data is poor. Much work is needed to improve model robustness.
Interpretability: Like most large neural models, Mistral 7B is difficult to interpret. Attention weights provide some insight but full explainability remains a challenge.
Fairness: Significant biases related to gender, race, and other attributes are observable in the model's outputs. Bias mitigation techniques were not thoroughly incorporated into the training process.
Generalization: Strong performance on hold-out test data indicates good generalization ability, but further testing is needed.
Customizability: The model can be fine-tuned to new tasks fairly easily, allowing customization. But adapting it to radically different domains requires extensive retraining.
Scalability: Efficiency gains from tensor processing units enable Mistral 7B to scale to handle heavy production workloads.
Compliance & Security: No major compliance or security issues have been identified yet, but more auditing is needed as deployment expands.
Environmental Impact: With an estimated carbon footprint of 200 tons CO2, Mistral 7B has high environmental costs. More work on efficiency and clean energy sources is recommended.
Human Evaluation: Limited human evaluation reveals good but imperfect conversational ability and domain knowledge. Expanded testing is advised.
User Experience: User testing shows reasonably fast and coherent responses, but some usability issues exist. Continued UX refinement is needed.
Cost: Mistral 7B was an expensive model to develop and deploy, costing millions of dollars. The payoff remains uncertain.
Reproducibility: Code, data, and model weights are publicly released enabling reproducibility.
Multilinguality: Only trained on English data, with no capabilities in other languages yet.
Mistral 7B shows promise but also clear areas for improvement as an AI assistant. Continued training, robustness enhancements, debiasing, and efficiency gains are recommended to improve its viability for production deployment. With refinement, it has the potential to become a versatile and valuable AI model.
Last updated