feat: implenent basic SFT pipeline based on synthetic data generator

[burtenshaw]

This is just an idea for how to make distilabel crazy-user-friendly. In this example, an SFT pipeline is abstracted away to make it easy to use. Like this:

from distilabel.presets import InstructionResponsePipeline

pipeline = InstructionResponsePipeline()

distiset = pipeline.run()

Maybe we could do this for other core tasks like: DPO, classification, retrieval

[codspeed-hq]

CodSpeed Performance Report

Merging #1059 will not alter performance

_{Comparing feat/pipeline-usecase-abstraction (79a13d5) with main (844165f)}

🎉 Hooray! pytest-codspeed just leveled up to 3.0.0!

A heads-up, this is a breaking change and it might affect your current performance baseline a bit. But here's the exciting part - it's packed with new, cool features and promises improved result stability 🥳!
Curious about what's new? Visit our releases page to delve into all the awesome details about this new version.

Summary

✅ 1 untouched benchmarks

[davidberenstein1957]

@burtenshaw what pipelines do you intend to create?

• SFT
• DPO
• Classification

Should we also have individual components like completion generation (if you only want an additional completion), instruction generation (which can be done with magpie and inheriting with only_instruction=True), preference (without generating instruction/completions from scratch)?

I can imagine people doing these things in steps and reviewing the quality of the generations in between before racking up costs with a complete pipeline.

[burtenshaw]

@burtenshaw what pipelines do you intend to create?

• SFT
• DPO
• Classification

@burtenshaw In this PR, I plan to just implement SFT and share that. But I think we could move on to DPO and classification straight after.

Should we also have individual components like completion generation (if you only want an additional completion), instruction generation (which can be done with magpie and inheriting with only_instruction=True), preference (without generating instruction/completions from scratch)?

I can imagine people doing these things in steps and reviewing the quality of the generations in between before racking up costs with a complete pipeline.

I agree that this is useful to more basic users. But this is also potentially complex. I don't want to end up abstracting the Magpie API in multiple ways, because the user could just use a proper pipeline. I would prefer to get feedback first.

[davidberenstein1957]

@burtenshaw what pipelines do you intend to create?

• SFT
• DPO
• Classification

@burtenshaw In this PR, I plan to just implement SFT and share that. But I think we could move on to DPO and classification straight after.

Should we also have individual components like completion generation (if you only want an additional completion), instruction generation (which can be done with magpie and inheriting with only_instruction=True), preference (without generating instruction/completions from scratch)?
I can imagine people doing these things in steps and reviewing the quality of the generations in between before racking up costs with a complete pipeline.

I agree that this is useful to more basic users. But this is also potentially complex. I don't want to end up abstracting the Magpie API in multiple ways, because the user could just use a proper pipeline. I would prefer to get feedback first.

Sure, sounds fair, I just wanted to see what your thoughts were surrounding the features.

[davidberenstein1957]

will you also add this to CLI commands?
something like the following would be cool
distilabel pipelines sft --num-rows 100 --n-turns 1

[burtenshaw]

That's nice, but we should do it another PR.

[plaguss]

I like the idea! Just let some small notes. If you can test it and generate a sample dataset that we can link in the docs that would be perfect

[burtenshaw]

@sdiazlor this is the abstraction for SFT in distilabel.

[burtenshaw]

I like the idea! Just let some small notes. If you can test it and generate a sample dataset that we can link in the docs that would be perfect

Here's a dataset: https://huggingface.co/datasets/argilla/distilabel-sft-easy