Add hard-negative flag to include similar challenging negatives on triplets

[plaguss]

Description

Adds a flag to GenerateSentencePair to generate hard-negatives that are more challenging to distinguish from the positives. It applies only when triplet=True. The distinction between the prompts is the following:

while the negative sentence
- can use similar words but must not be related to the anchor sentence.
+ is a 'hard negative' that meets the following criteria:\n
+ - Uses similar keywords or phrases as the anchor sentence\n
+ - Has a similar grammatical structure or syntax\n
+ - Is not related to the anchor sentence, but could be mistaken for it\n
+ Try to create a negative sentence that would be challenging for a model to distinguish 
+ from the positive sentence
You must output only two new sections: `## Positive` and `## Negative`."

By default, the standard hard_negative=False will be used.

from distilabel.steps.tasks import GenerateSentencePair
from distilabel.llms import InferenceEndpointsLLM

generate_sentence_pair = GenerateSentencePair(
    triplet=True, # `False` to generate only positive
    action="query",
    context="Argilla is an open-source data curation platform for LLMs.",
    hard_negative=True,
    llm=InferenceEndpointsLLM(model_id="meta-llama/Meta-Llama-3.1-70B-Instruct"),
)

Closes #789

[github-actions]

Documentation for this PR has been built. You can view it at: https://distilabel.argilla.io/pr-856/

[codspeed-hq]

CodSpeed Performance Report

Merging #856 will not alter performance

_{Comparing sentence-pair-hard-negatives (e5e3f27) with develop (4e1f2bc)}

Summary

✅ 1 untouched benchmarks