I was testing translategemma's 12B version with one of the banking domain use cases and the responses I got for some of the numerical data (written in words) were completely absurd. The source lang was "en" and target was "hi"(hindi). Here is an example of one of the responses
{
"id": 26,
"en": "The principal amount of fifty thousand rupees is to be repaid over thirty-six months.",
"hi_ref": "पचास हज़ार रुपये की मूल राशि छत्तीस महीनों में चुकानी होगी।",
"hi_opt": "पच्चीस हजार रुपये की मूल राशि को तीस महीने की अवधि में चुकाया जाना है।"
}
The reference mentions "50 thousand" but the output says "25 thousand" (in hindi). Has anyone faced similar issues and how can it be resolved?
I was testing translategemma's 12B version with one of the banking domain use cases and the responses I got for some of the numerical data (written in words) were completely absurd. The source lang was "en" and target was "hi"(hindi). Here is an example of one of the responses
{
"id": 26,
"en": "The principal amount of fifty thousand rupees is to be repaid over thirty-six months.",
"hi_ref": "पचास हज़ार रुपये की मूल राशि छत्तीस महीनों में चुकानी होगी।",
"hi_opt": "पच्चीस हजार रुपये की मूल राशि को तीस महीने की अवधि में चुकाया जाना है।"
}
The reference mentions "50 thousand" but the output says "25 thousand" (in hindi). Has anyone faced similar issues and how can it be resolved?