[Feature Request] Add EvolInstruct methods to camel/datagen

[cat-searcher-bot]

Required prerequisites

• I have searched the Issue Tracker and Discussions that this hasn't already been reported. (+1 or comment there if it has.)
• Consider asking first in a Discussion.

Motivation

enhancing the current camel/datagen pipeline with new instruction generation methods.

Solution

-- will make a pull request shortly.

Alternatives

There are many other interesting prompt synthesis works that we may implement, e.g., Magpie, EvoPrompt, etc.

Additional context

Evol-Instruct is a simple-to-use method for generating new prompts. basically, given an existing prompt p, it applies a re-write instruction r on it, and obtain a new prompt p' ~ LLM(⋅| p, r). in general, the re-write instructions include in-depth and in-breadth re-writing. the process can be recursive.

• the initial paper: WizardLM: Empowering Large Language Models to Follow Complex Instructions.
• existing implementations: WizardLM/EvolInstruct, distilabel/EvolInstruct.
• reference usage in RL: Evolving Alignment via Asymmetric Self-Play: Scalable Reinforcement Post-Training Beyond Static Human Prompts.