Phi 3 Mini 4K Instruct is an open-source language model by microsoft. Features: 3.8b LLM, VRAM: 7.7GB, Context: 4K, License: mit, Instruction-Based, LLM Explorer Score: 0.44, ELO: 1081, Arc: 63, HellaSwag: 80.6, MMLU: 69.1, GSM8K: 74.5.
Phi 3 Mini 4K Instruct Benchmarks
nn.n% — How the model compares to the reference models: Anthropic Sonnet 3.5 ("so35"), GPT-4o ("gpt4o") or GPT-4 ("gpt4").
Phi 3 Mini 4K Instruct Parameters and Internals
| Model Type | | text generation, language model |
|
| Use Cases |
| Areas: | | Research, Commercial applications |
|
| Applications: | | General purpose AI systems, Computationally constrained environments |
|
| Primary Use Cases: | | Memory/computational constraint scenarios, Latency bound applications, Tasks requiring mathematical and logical reasoning |
|
| Limitations: | | Limited by language/data representation bias, Requires additional debiasing techniques for high-risk use cases |
|
| Considerations: | | Evaluate performance and mitigate for safety and accuracy. |
|
|
| Additional Notes | | The model's performance improves when integrated with retrieval systems for external knowledge. |
|
| Supported Languages | | en (High proficiency), fr (Moderate proficiency) |
|
| Training Details |
| Data Sources: | | Publicly available documents, High-quality educational data, Newly created synthetic data, Chat format supervised data |
|
| Data Volume: | |
| Methodology: | | Supervised fine-tuning and Direct Preference Optimization |
|
| Context Length: | |
| Training Time: | |
| Hardware Used: | |
| Model Architecture: | | Dense decoder-only Transformer model |
|
|
| Safety Evaluation |
| Methodologies: | | Supervised fine-tuning, Direct Preference Optimization |
|
| Findings: | | Strong reasoning capabilities, Improved instruction following |
|
| Risk Categories: | |
| Ethical Considerations: | | Use responsibly and ensure compliance with laws. |
|
|
| Responsible Ai Considerations |
| Fairness: | | Address bias through training data selection and filtering. |
|
| Transparency: | | Encourage user feedback and continuous improvement. |
|
| Accountability: | | Developers responsible for outputs and compliance. |
|
| Mitigation Strategies: | | Use Retrieval Augmented Generation for grounding responses. |
|
|
| Input Output |
| Input Format: | |
| Accepted Modalities: | |
| Output Format: | |
| Performance Tips: | | Utilize chat format for best results. Consider prompt engineering for improved performance. |
|
|
| Release Notes |
| Version: | |
| Notes: | | Improved instruction following, structure output, and reasoning compared to the original release. |
|
|
|
Quantized Models of the Phi 3 Mini 4K Instruct
Best Alternatives to Phi 3 Mini 4K Instruct