Ensure HF compatibility with sentence-transformers/embedding-training-data

[jzombie]

https://huggingface.co/datasets/sentence-transformers/embedding-training-data/blob/main/README.md

The JSON objects can come in different formats:

• Pairs: ["text1", "text2"] - This is a positive pair that should be close in vector space.
• Triplets: ["anchor", "positive", "negative"] - This is a triplet: The positive text should be close to the anchor, while the negative text should be distant to the anchor.
• Sets: {"set": ["text1", "text2", ...]} A set of texts describing the same thing, e.g. different paraphrases of the same question, different captions for the same image. Any combination of the elements is considered as a positive pair.
• Query-Pairs: {"query": "text", "pos": ["text1", "text2", ...]} A query together with a set of positive texts. Can be formed to a pair ["query", "positive"] by randomly selecting a text from pos.
• Query-Triplets: {"query": "text", "pos": ["text1", "text2", ...], "neg": ["text1", "text2", ...]} A query together with a set of positive texts and negative texts. Can be formed to a triplet ["query", "positive", "negative"] by randomly selecting a text from pos and neg.

[jzombie]

Also debug:

• hf://shash42/forecast-news anchor=title positive=content
• hf://gbharti/finance-alpaca anchor=instruction positive=output

Edit: Fixes for these will land with: #157