Paper Summary: Zephyr: Direct Distillation of LM Alignment

Please note This post is mainly intended for my personal use. It is not peer-reviewed work and should not be taken as such.

Zephyr: Direct Distillation of LM Alignment Source

WHAT

Authors instruction-tune Mistral-7B vanilla by distillation: using DPO on open preference datasets and samples generated from previously aligned teacher models.

WHY

Because traditional distillation strategies are only good at transferring stylistic — not alignment capabilities.

HOW

Starting with Mistral-7B as the V0 model:

• 1) Run SFT on V0 using input/output pairs from the UltraChat dataset, generating model V1

• 3) Use inputs from UltraFeedback dataset and, for each input, feed it to intermediary models (Claude, Falcon, etc), generating multiple output variations for the same input.

• 4) For each input from step 3, feed all the output variations to the teacher model (GPT-4) and ask it to select the best one.

• 5) Use DPO to align model V1, using the best output for each input, as selected on step 4.¹

CLAIMS

• It's possible to transfer alignment capabilities from teacher models using the suggested approach.

• The DPO model overfits quickly with longer training.

• Zephyr-7B outperforms 70B models (such as Llama-chat-70B) on some benchmarks.

QUOTES

• "... without an initial SFT step ... models are not able to learn at all from feedback and perform terribly."
  • This is interesting. We can't jump to reward modeling without the initial SFT step.

EXTENDS/USES

• Mistral-7B

• Other aligned LLMs as teachers: Claude, Falcon, Llama, GPT-4.

NOTES

• Distillation appears to be the default term for extracting the capabilities of a "teacher" model into a simpler and cheaper "student" model. Apparently it was introduced by Hinton et al 2015.

• Zephyr-7B was fully optimized for Helpfulness only.

1: More precisely, DPO is optimized using the best response to each each, but contrasting it to a randomly chosen response. It doesn't classify response, it ranks them