Phi 1.5 Instruct: an instruction following Phi 1.5 model that has undergone SFT and DPO

#89
by rasyosef - opened

Phi-1_5-Instruct-v0.1

https://huggingface.co/rasyosef/Phi-1_5-Instruct-v0.1

Phi-1_5-Instruct is an instruction following version of Microsoft's Phi 1.5 base model that has underwent a post-training process that incorporates both supervised fine-tuning and direct preference optimization. I used the trl library and a single A100 40GB GPU during both the SFT and DPO steps.

How to use

Chat Format

Given the nature of the training data, the Phi-1.5 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

This model outperforms HuggingFace's SmolLM-1.7B-Instruct and the TinyLlama-1.1B-Chat-v1.0 models on all five of the following 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.
Model Size (# params) IFEval GSM8K MMLU TruthfulQA Winogrande
Phi-1_5-Instruct-v0.1 1.4B 26.71 41.78 39.72 47.9 70.4
SmolLM-1.7B-Instruct 1.7B 24.21 3.45 23.57 47.38 63.61
TinyLlama-1.1B-Chat-v1.0 1.1B 21.23 0 24.03 39.14 61.01
phi-1_5 1.4B 20.51 31.73 42.48 40.86 71.74

Demo

You can use this hugging face space to interact with the chat model
https://huggingface.co/spaces/rasyosef/phi-1_5-chat

Sign up or log in to comment