This model literally beats all the other models in the needle in the haystack challenge
I built 4 dataset (quite short but long enough)
Their mission is to do needle in the haystack in the following context :
Multilingual : easy match (multilingual_easy)
Multilingual : hard match (multilingual_subtle)
Crosslingual : easy match (crosslingual_easy)
Crosslingual : hard match (crosslingual_subtle)
Dataset sample (not complete:
dataset_multilingual_easy
{
"question": "Qui est l'ancien propio du Rablugorda ?",
"needles": ["Le précédent propriétaire du Rablugorda était un fermier espagnol."]
},
{
"question": "¿Cuándo nació Peter Donkey?",
"needles": ["Peter Donkey nació en noviembre de 1996"]
},
{
"question": "Wie hoch ist der Mount Everest?",
"needles": ["Der Mount Everest ist 8.848 Meter hoch."]
},
{
"question": "Who invented the first modern computer?",
"needles": ["Turing was instrumental in creating the foundational concepts of modern computing."]
},
...
dataset_multilingual_subtle
{
"question": "Когда Питер съел фрукт?",
"needles": ["После интенсивной тренировки он восстановил силы натуральным способом."]
},
{
"question": "Ce qu'a fait le criminel pour aller en prison?",
"needles": ["Un moment d'égarement dans l'épicerie du quartier a bouleversé sa vie entière."]
},
{
"question": "Qual foi a primeira descoberta do cientista?",
"needles": ["Uma observação acidental no laboratório mudou todo o rumo da pesquisa."]
},
...
dataset_crosslingual_easy
{
"question": "Qui est l'ancien propio du Rablugorda ?",
"needles": [
"Little did he know, the owner of the Rablugorda was none other than the king himself.",
"The previous owner of the Rablugorda was a spanish farmer.",
"El propietario anterior del Rablugorda era un granjero espa\u00f1ol.",
"Die vorherige Besitzerin des Rablugorda war eine spanische B\u00e4uerin."
]
},
{
"question": "When was Peter Donkey Born ?",
"needles": [
"Peter Donkey est n\u00e9 en novembre 1996",
"Peter Donkey ese nacio en 1996",
"Peter Donkey wurde im November 1996 geboren"
]
},
{
"question": "Wie hoch ist der Mount Everest?",
"needles": [
"Mount Everest measures 8,848 meters above sea level.",
"The tallest mountain is 8,848 meters high.",
"La montagne la plus haute mesure 8 848 m\u00e8tres, c'est l'Everest.",
"La monta\u00f1a m\u00e1s alta mide 8,848 metros."
]
},
dataset_crosslingual_subtle
{
"question": "Qual foi a primeira descoberta do cientista?",
"needles": [
"A breakthrough that challenged existing paradigms.",
"Une r\u00e9volution dans la compr\u00e9hension scientifique.",
"Ein Moment, der die Forschung f\u00fcr immer ver\u00e4nderte.",
"The insight that opened new paths of understanding."
]
},
{
"question": "Wie lernte der Musiker das Komponieren?",
"needles": [
"Through passion and dedication to his craft.",
"Un parcours de transformation artistique.",
"La b\u00fasqueda incesante de la perfecci\u00f3n mel\u00f3dica.",
"A journey of emotional and technical mastery."
]
},
{
"question": "Onde o arquiteto encontrou sua inspira\u00e7\u00e3o?",
"needles": [
"In the intersection of nature and human design.",
"Un dialogue entre tradition et innovation.",
"La fusi\u00f3n perfecta de forma y funci\u00f3n.",
"Die Harmonie zwischen Raum und Menschlichkeit."
]
},
...
Thank you for your contribution and sharing. Very interesting work!
By the way, is there a publicly available paper or blog post for a more in-depth understanding of these results?
I plan to release a github repository with the available code, xlsx and datasets, the datasets are quite short, only like ~20-40 entries for now buit it's enough to see that some embeddings model and some rerankers are capable or not capable of acheiving some tasks.
As for a paper release I'll need to discuss of it with a pair more in depth
I plan to release a github repository with the available code, xlsx and datasets, the datasets are quite short, only like ~20-40 entries for now buit it's enough to see that some embeddings model and some rerankers are capable or not capable of acheiving some tasks.
Looking forward to your interesting work. Feel free to discuss and exchange ideas at any time.