|
--- |
|
license: cc-by-nc-4.0 |
|
language: |
|
- tr |
|
--- |
|
|
|
# Model Card for Model ID |
|
|
|
<!-- Provide a quick summary of what the model is/does. --> |
|
|
|
gemma-2b-tr fine-tuned with Turkish Instruction-Response pairs. |
|
|
|
## Model Details |
|
|
|
### Model Description |
|
|
|
<!-- Provide a longer summary of what this model is. --> |
|
|
|
- **Language(s) (NLP):** Turkish, English |
|
- **License:** Creative Commons Attribution Non Commercial 4.0 |
|
- **Finetuned from model [optional]:** gemma-2b-tr (https://huggingface.co/Metin/gemma-2b-tr) |
|
|
|
|
|
## Uses |
|
|
|
The model is designed for Turkish instruction following and question answering. Its current response quality is limited, likely due to the small instruction set and model size. It is not recommended for real-world applications at this stage. |
|
|
|
## Restrictions |
|
|
|
Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms |
|
Please refer to the gemma use restrictions before start using the model. |
|
https://ai.google.dev/gemma/terms#3.2-use |
|
|
|
## How to Get Started with the Model |
|
|
|
```Python |
|
from transformers import AutoTokenizer, AutoModelForCausalLM |
|
|
|
tokenizer = AutoTokenizer.from_pretrained("Metin/gemma-2b-tr-inst") |
|
model = AutoModelForCausalLM.from_pretrained("Metin/gemma-2b-tr-inst") |
|
|
|
system_prompt = "You are a helpful assistant. Always reply in Turkish." |
|
instruction = "Ankara hangi ülkenin başkentidir?" |
|
prompt = f"{system_prompt} [INST] {instruction} [/INST]" |
|
input_ids = tokenizer(prompt, return_tensors="pt") |
|
|
|
outputs = model.generate(**input_ids) |
|
print(tokenizer.decode(outputs[0])) |
|
``` |
|
|
|
As it can be seen from the above example instructions should be framed within the following structure: |
|
|
|
SYSTEM_PROMPT [INST] \<Your instruction here\> [/INST] |
|
|
|
## Training Details |
|
|
|
### Training Data |
|
|
|
- Dataset: Turkish instructions from the Aya dataset (https://huggingface.co/datasets/CohereForAI/aya_dataset) |
|
- Dataset size: ~550K Token or ~5K instruction-response pair. |
|
|
|
### Training Procedure |
|
|
|
#### Training Hyperparameters |
|
|
|
- **Adapter:** QLoRA |
|
- **Epochs:** 1 |
|
- **Context length:** 1024 |
|
- **LoRA Rank:** 32 |
|
- **LoRA Alpha:** 32 |
|
- **LoRA Dropout:** 0.05 |
