Fine tuning precision task of transliteration on Gemma-2b-it model

#113
by grishi911991 - opened

Facing issues with finetuning gemma-2b-it model for transliteration task.

I want create a lora for transliteration task from Roman to Devnagri and vice versa. But even after multiple combination of

  1. Lora rank
  2. Lora decay
  3. Modules
  4. Lora drop out rate
    etc

When I inferred the results for the same Lora using oogabooga, it is not generated desired result, rather it just repeated user content. Even though I have used similar parameters for training lora for same task on gemma-9b model, it is working fine.

Can someone help here or have some thoughts.

Sign up or log in to comment