README.md

---
license: apache-2.0
base_model: meta-llama/Llama-2-13b-hf
tags:
  - generated_from_trainer
  - llama
  - lora
  - adapters
datasets:
  - yhavinga/mc4_nl_cleaned
language:
  - nl
model-index:
  - name: llama2-13b-ft-mc4_nl_cleaned_tiny
    results: []
---
    

# llama2-13b-ft-mc4_nl_cleaned_tiny

This model is a fine-tuned version of [meta-llama/Llama-2-13b-hf](https://huggingface.co/meta-llama/Llama-2-13b-hf)
on the [yhavinga/mc4_nl_cleaned](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned/viewer/tiny/train) dataset (`tiny` partition) on a context of 4096 tokens.
See the original [meta-llama/Llama-2-13b-hf](https://huggingface.co/meta-llama/Llama-2-13b-hf) for more information, intended use, and biases.

## Intended uses & limitations

While Llama 2 already contains some proficiency in Dutch, this finetune is intended to improve the fluency of Dutch (not increase its knowledge). It is therefore
intended as a generative model for Dutch language. The biases, shortcomings and intended uses are otherwise the same as those of
the [original model]((https://huggingface.co/meta-llama/Llama-2-13b-hf)). The model can be used for generative tasks or finetuned further on other tasks
such as summarization, adaptation, instruction or chat finetuning.

## Training and evaluation data

Trained on the [yhavinga/mc4_nl_cleaned](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned/viewer/tiny/train) dataset (`tiny` partition) for one epoch. The canonical 
validation split was not used but instead 5% of `train` was used as validation.

## Training procedure

Trained with LoRA targetting `["q_proj", "v_proj"]` in 4 bit and merged before upload. Trained with Flash Attention as borrowed from
[here](https://github.com/philschmid/deep-learning-pytorch-huggingface/blob/main/training/utils/llama_patch.py).

The adapters are in the `adapters` branch.

Initial training investigation on the Tier-1 HPC of [Vlaams Supercomputer Centrum (VSC)](https://www.vscentrum.be/) and training on our own server of 4x 3090s.


### Training hyperparameters

The following hyperparameters were used during training in the HPC investigation:
- learning_rate: 0.0003
- train_batch_size: 12
- eval_batch_size: 12
- seed: 42
- distributed_type: multi-GPU
- num_devices: 16
- gradient_accumulation_steps: 6
- total_train_batch_size: 1152
- total_eval_batch_size: 192
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.03
- num_epochs: 1

### Training results

| Training Loss | Epoch | Step | Validation Loss |
|:-------------:|:-----:|:----:|:---------------:|
| 1.8784        | 0.09  | 90   | 1.8820          |
| 1.8344        | 0.19  | 180  | 1.8542          |
| 1.8351        | 0.28  | 270  | 1.8355          |
| 1.8206        | 0.37  | 360  | 1.8212          |
| 1.8021        | 0.47  | 450  | 1.8088          |
| 1.8102        | 0.56  | 540  | 1.7982          |
| 1.7991        | 0.65  | 630  | 1.7890          |
| 1.7788        | 0.74  | 720  | 1.7811          |
| 1.7915        | 0.84  | 810  | 1.7742          |
| 1.7715        | 0.93  | 900  | 1.7676          |


### Framework versions

- Transformers 4.31.0.dev0
- Pytorch 2.0.1+cu117
- Datasets 2.13.1
- Tokenizers 0.13.3

# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_BramVanroy__llama2-13b-ft-mc4_nl_cleaned_tiny)

| Metric                | Value                     |
|-----------------------|---------------------------|
| Avg.                  | 46.81   |
| ARC (25-shot)         | 59.3          |
| HellaSwag (10-shot)   | 82.04    |
| MMLU (5-shot)         | 54.67         |
| TruthfulQA (0-shot)   | 38.03   |
| Winogrande (5-shot)   | 77.27   |
| GSM8K (5-shot)        | 10.31        |
| DROP (3-shot)         | 6.08         |