Hugging Face Overview
rasyosef/Phi-1_5-Instruct-v0.1: Instruction-Tuned Small Language Model
This model is a 1.4 billion parameter Transformer, developed by rasyosef, based on the Microsoft Phi-1.5 architecture. It has undergone a post-training process combining supervised fine-tuning (SFT) and direct preference optimization (DPO) to enhance its instruction-following capabilities. The SFT phase utilized 128,000 instruction-response pairs from the teknium/OpenHermes-2.5 dataset, while DPO leveraged a combination of preference datasets including HuggingFaceH4/ultrafeedback_binarized and argilla/distilabel-intel-orca-dpo-pairs.
Key Capabilities & Performance
- Instruction Following: Excels at adhering to explicit instructions, as measured by IFEval.
- Mathematical Reasoning: Demonstrates strong performance on grade school math problems (GSM8k).
- Multitask Accuracy: Achieves competitive results across 57 diverse tasks in MMLU.
- Commonsense Reasoning: Performs well on the Winogrande benchmark.
- Benchmark Superiority: Outperforms HuggingFace's SmolLM-1.7B-Instruct and TinyLlama-1.1B-Chat-v1.0 across IFEval, GSM8K, MMLU, TruthfulQA, and Winogrande benchmarks.
Ideal Use Cases
- Instruction-based tasks: Suited for applications requiring precise adherence to user prompts and formatting.
- Reasoning tasks: Effective for common sense, logical, and mathematical problem-solving in a small model footprint.
- Resource-constrained environments: Its 1.4B parameter size makes it efficient for deployment where larger models are impractical.