phi-2-instruct-v0.1

https://huggingface.co/rasyosef/phi-2-instruct-v0.1

The model has underwent a post-training process that incorporates both supervised fine-tuning and direct preference optimization for instruction following. I used the trl library and a single A100 40GB GPU during both the SFT and DPO steps.

• Supervised Fine-Tuning

  • SFT Model: phi-2-sft
  • Used 128,000 instruction, response pairs from the teknium/OpenHermes-2.5 dataset

• Direct Preference Optimization (DPO)

  • LoRA Adapter: phi-2-dpo
  • Used a combination of the following preference datasets
    • HuggingFaceH4/ultrafeedback_binarized
    • argilla/distilabel-intel-orca-dpo-pairs
    • argilla/distilabel-math-preference-dpo

How to use

Chat Format

Given the nature of the training data, the phi-2 instruct model is best suited for prompts using the chat format as follows.
You can provide the prompt as a question with a generic template as follow:

<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
Question?<|im_end|>
<|im_start|>assistant

Benchmarks

These benchmarks were run using EleutherAI's lm-evaluation-harness

• IFEval (Instruction Following Evaluation): IFEval is a fairly interesting dataset that tests the capability of models to clearly follow explicit instructions, such as “include keyword x” or “use format y”. The models are tested on their ability to strictly follow formatting instructions rather than the actual contents generated, allowing strict and rigorous metrics to be used.
• GSM8k (5-shot): diverse grade school math word problems to measure a model's ability to solve multi-step mathematical reasoning problems.
• MMLU (5-shot) - a test to measure a text model's multitask accuracy. The test covers 57 tasks including elementary mathematics, US history, computer science, law, and more.
• TruthfulQA - a test to measure a model's propensity to reproduce falsehoods commonly found online. Note: TruthfulQA is technically a 6-shot task in the Harness because each example is prepended with 6 Q/A pairs, even in the 0-shot setting.
• Winogrande (5-shot) - an adversarial and difficult Winograd benchmark at scale, for commonsense reasoning.