README.md

---
language:
- nl
license: cc-by-nc-4.0
library_name: peft
tags:
- llama
- alpaca
- Transformers
- text-generation-inference
datasets:
- BramVanroy/alpaca-cleaned-dutch
inference: false
base_model: openlm-research/open_llama_13b
pipeline_tag: text-generation
---

# open_llama_13b_alpaca_clean_dutch_qlora

## Model description

This adapter model is a fine-tuned version of [openlm-research/open_llama_13b](https://huggingface.co/openlm-research/open_llama_13b). 
Finetuning was performed on the Dutch [BramVanroy/alpaca-cleaned-dutch](https://www.huggingface.co/datasets/BramVanroy/alpaca-cleaned-dutch) dataset which contains 52K of records with instruction following-data translated from English to Dutch.

See [openlm-research/open_llama_13b](https://huggingface.co/openlm-research/open_llama_13b) for all information about the base model.

## Model usage

A basic example of how to use the finetuned model.

```
import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "robinsmits/open_llama_13b_alpaca_clean_dutch_qlora"

tokenizer =  AutoTokenizer.from_pretrained(model_name, use_fast = False, add_eos_token = True)

config = PeftConfig.from_pretrained(model_name)

model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, load_in_8bit = True, device_map = "auto")
model = PeftModel.from_pretrained(model, model_name)

prompt = "### Instructie:\nWat zijn de drie belangrijkste softwareonderdelen die worden gebruikt bij webontwikkeling?\n\n### Antwoord:\n" 

inputs = tokenizer(prompt, return_tensors = "pt", truncation = True).input_ids.cuda()
sample = model.generate(input_ids = inputs, max_new_tokens = 512, num_beams = 2, early_stopping = True, eos_token_id = tokenizer.eos_token_id)
output = tokenizer.decode(sample[0], skip_special_tokens = True)

print(output.split(prompt)[1])
```

The prompt and generated output for the above mentioned example is similar to the output shown below.

```
### Instructie:
Wat zijn de drie belangrijkste softwareonderdelen die worden gebruikt bij webontwikkeling?

### Antwoord:
 Optional
1. HTML
2. CSS
3. JavaScript
```

For more extensive usage and a lot of generated samples (both good and bad samples) see the following [Inference Notebook](https://github.com/RobinSmits/Dutch-LLMs/blob/main/Open_Llama_13B_Alpaca_Clean_Dutch_Inference.ipynb)

## Intended uses & limitations

The open_llama_13b model was primarily trained on the English language. Part of the dataset was a Wikipedia dump containing pages in 20 languages.
Dutch was one of those languages. Given the size of the total dataset and the wikipedia part the Dutch language was very likely less than 0.5% of the total data. 

The generated output and performance of this model for the Dutch language is very likely not always comparable to the various Open-Llama models that have been finetuned on English Alpaca datasets.

The primary intention of this model is to explore and research the use of the Dutch language in combination with an Open LLM model.

## Training and evaluation data

This model was trained on the [BramVanroy/alpaca-cleaned-dutch](https://www.huggingface.co/datasets/BramVanroy/alpaca-cleaned-dutch) dataset.

Based on the dataset license only Non-Commercial use is allowed. Commercial use is strictly forbidden.

```
@misc{vanroy2023language,
      title={Language Resources for {Dutch} Large Language Modelling}, 
      author={Bram Vanroy},
      year={2023},
      eprint={2312.12852},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```

## Training procedure

This model was finetuned with a QLoRA setup on a Google Colab A100 GPU in about 11 hours.

The notebook used for training can be found here: [Training Notebook](https://github.com/RobinSmits/Dutch-LLMs/blob/main/Open_Llama_13B_Alpaca_Clean_Dutch_Qlora.ipynb)

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0002
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 64
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 64
- training_steps: 1536

The following `bitsandbytes` quantization config was used during training:
- load_in_8bit: False
- load_in_4bit: True
- llm_int8_threshold: 6.0
- llm_int8_skip_modules: None
- llm_int8_enable_fp32_cpu_offload: False
- llm_int8_has_fp16_weight: False
- bnb_4bit_quant_type: nf4
- bnb_4bit_use_double_quant: True
- bnb_4bit_compute_dtype: bfloat16

### Training results

| Training Loss | Epoch | Step | Validation Loss |
|:-------------:|:-----:|:----:|:---------------:|
| 1.0439        | 1.0   | 768  | 1.0496          |
| 0.9431        | 2.0   | 1536 | 0.9879          |

### Framework versions

- Transformers 4.30.2
- Pytorch 2.0.1+cu118
- Datasets 2.13.1
- Tokenizers 0.13.3
- PEFT 0.4.0.dev0