Phi 1 5 is an open-source language model by microsoft. Features: 1.4b LLM, VRAM: 2.8GB, Context: 2K, License: mit, LLM Explorer Score: 0.22, Arc: 55.4, HellaSwag: 57.1, MMLU: 44.3, GSM8K: 32.5.
Phi 1 5 Parameters and Internals
| Model Type | | Transformer, text generation, NLP, code |
|
| Use Cases |
| Areas: | |
| Applications: | | text generation, code generation |
|
| Primary Use Cases: | | poem writing, email drafting, story creation, text summarization, Python code writing |
|
| Limitations: | | Potential to generate harmful content, Generate inaccurate code and facts, Unreliable responses to instruction, Limited scope for code |
|
| Considerations: | | Users should be cautious and critically evaluate outputs. |
|
|
| Additional Notes | | Phi-1.5-generated text/code should be treated as a starting point. Users should verify API uses manually where uncommon packages are involved. |
|
| Supported Languages | |
| Training Details |
| Data Sources: | | same data sources as phi-1, various NLP synthetic texts |
|
| Data Volume: | |
| Training Time: | |
| Hardware Used: | |
| Model Architecture: | | Transformer-based with next-word prediction objective |
|
|
| Responsible Ai Considerations |
| Transparency: | | The model has not undergone instruction fine-tuning. |
|
| Mitigation Strategies: | | Model is intended for research to help develop methods to reduce toxicity directly after pretraining. |
|
|
| Input Output |
| Input Format: | | QA format, Chat format, Code format |
|
| Accepted Modalities: | |
| Output Format: | |
| Performance Tips: | | Users should update to `transformers` version 4.37.0 or higher. |
|
|
Quantized Models of the Phi 1 5
Best Alternatives to Phi 1 5