17 lines
1.0 KiB
Markdown
17 lines
1.0 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
datasets:
|
||
|
|
- ihalage/sinhala-instruction-finetune-large
|
||
|
|
language:
|
||
|
|
- si
|
||
|
|
- en
|
||
|
|
---
|
||
|
|
[](https://github.com/ihalage/llama3-sinhala)
|
||
|
|
|
||
|
|
LLaMA3 (8B) model instruction finetuned to understand and respond in Sinhala language. `meta-llama/Meta-Llama-3-8B-Instruct` is finetuned on a reletively large dataset in
|
||
|
|
Sinhala compiled by translating English datasets such as ELI5 and Alpaca. The dataset in hosted in Hugging Face Datasets hub [(`sinhala-instruction-finetune-large`)](https://huggingface.co/datasets/ihalage/sinhala-instruction-finetune-large)
|
||
|
|
|
||
|
|
The original model is 4-bit quantized and finetuned with a causal language modelling (CLM) objective by adding LoRA adapters with a rank of 16 and a scaling factor of 32.
|
||
|
|
|
||
|
|
The finetuned `llama3-sinhala` model generates better responses in Sinhala compared to the original instruction finetuned model released by Meta.
|
||
|
|
See the github repo [llama3-sinhala](https://github.com/ihalage/llama3-sinhala) for more details.
|