|
|
--- |
|
|
library_name: peft |
|
|
model_name: lora_1B_TR |
|
|
tags: |
|
|
- meta-llama/Llama-3.2-1B-Instruct |
|
|
- lora |
|
|
- sft |
|
|
- transformers |
|
|
- trl |
|
|
- unsloth |
|
|
licence: license |
|
|
pipeline_tag: text-generation |
|
|
base_model: meta-llama/Llama-3.2-1B-Instruct |
|
|
datasets: |
|
|
- kadirnar/combined-turkish-datasets-v5 |
|
|
language: |
|
|
- tr |
|
|
- en |
|
|
--- |
|
|
|
|
|
|
|
|
|
|
|
# Model Card for Lora_TR_1B |
|
|
|
|
|
This is a Lora Adaptor of 'meta-llama/Llama-3.2-1B-Instruct'. |
|
|
The main goal of this adapter is to obtain an Llama who speaks Turkish better. |
|
|
>(r=32, lora_alpha=64, lora_dropout=0.005) |
|
|
|
|
|
## Quick start |
|
|
|
|
|
```python |
|
|
from unsloth import FastLanguageModel |
|
|
from peft import PeftModel |
|
|
from transformers import AutoTokenizer |
|
|
|
|
|
BASE = "meta-llama/Llama-3.2-1B-Instruct" |
|
|
ADAPTER = "Codex07/Lora_1B_TR" |
|
|
|
|
|
# Load Model |
|
|
model, tok = FastLanguageModel.from_pretrained( |
|
|
model_name=BASE, max_seq_length=2048, load_in_4bit=False, dtype=None, device_map="auto" |
|
|
) |
|
|
# Load Adaptor |
|
|
model = PeftModel.from_pretrained(model, ADAPTER) # adapter’ı Unsloth modeline tak |
|
|
FastLanguageModel.for_inference(model) |
|
|
|
|
|
# Test |
|
|
messages = [ |
|
|
{"role":"system","content":"You are AI assistant. Give user answers"},# Sen bir Yapay Zeka Asistanısısın. kullanıcıdan gelen sorulara resmi cevap ver. |
|
|
{"role":"user","content":"Selam!"} |
|
|
] |
|
|
prompt = tok.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device) |
|
|
out = model.generate(prompt, max_new_tokens=2048) |
|
|
print(tok.decode(out[0, prompt.shape[-1]:], skip_special_tokens=True)) |
|
|
``` |
|
|
|
|
|
## Training procedure |
|
|
Half of 'kadirnar/combined-turkish-datasets-v5' Turkish dataset used. |
|
|
Dataset divided into chunks by size 65k. |
|
|
|
|
|
```bibtex |
|
|
1> 2:50:33 / 2.746500 -> 1.771400 / 5.1.0 |
|
|
2> 3:00:00 / 1.7 -> 1.7 / 5.1.1 |
|
|
3> 2:18:19 / 1.859100 -> 1.474300 / 5.1.2 |
|
|
4> 3:15:13 / 1.421800 -> 1.122000 / 5.1.3 |
|
|
5> 2:50:00 / 1.746600 -> 1.629600 / 5.1.0 |
|
|
6> 2:44:46 / 1.745000 -> 1.653300 / 5.1.1 |
|
|
7> 2:07:00 / 1.478200 -> 1.357400 / 5.1.2 |
|
|
8> 3:11:54 / 1.174700 -> 1.046100 / 5.1.3 |
|
|
9> 3:12:39 / 1.117600 -> 0.796700 / 5.2.0 |
|
|
10>1:00:57 / 2.217400 -> 1.741400 / 5.2.1 |
|
|
11>1:30:04 / 2.919900 -> 2.534300 / 5.2.2 |
|
|
12>1:30:05 / 2.534300 -> 2.320100 / 5.2.2 |
|
|
``` |
|
|
|
|
|
This model was trained with SFT. |
|
|
|
|
|
### Framework versions |
|
|
|
|
|
- PEFT 0.17.1 |
|
|
- TRL: 0.23.0 |
|
|
- Transformers: 4.56.2 |
|
|
- Pytorch: 2.8.0 |
|
|
- Datasets: 4.3.0 |
|
|
- Tokenizers: 0.22.1 |
|
|
|
|
|
## Citations |
|
|
|
|
|
Cite TRL as: |
|
|
```bibtex |
|
|
@misc{vonwerra2022trl, |
|
|
title = {{TRL: Transformer Reinforcement Learning}}, |
|
|
author = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec}, |
|
|
year = 2020, |
|
|
journal = {GitHub repository}, |
|
|
publisher = {GitHub}, |
|
|
howpublished = {\url{https://github.com/huggingface/trl}} |
|
|
} |
|
|
``` |
