File size: 10,902 Bytes
c9802b9 638c4cb c9802b9 638c4cb c9802b9 638c4cb c9802b9 638c4cb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 |
---
language:
- nl
license: llama3
---
<p align="center" style="margin:0;padding:0">
<img src="./chocollama_logo.png" alt="ChocoLlama logo" width="500" style="margin-left:'auto' margin-right:'auto' display:'block'"/>
</p>
<div style="margin:auto; text-align:center">
<h1 style="margin-bottom: 0">ChocoLlama</h1>
<em>A Llama-2/3-based family of Dutch language models</em>
</div>
## Llama-3-ChocoLlama-8B-base: Getting Started
We here present **Llama-3-ChocoLlama-8B-base**, a language-adapted version of Meta's Llama-3-8b, fine-tuned on 17B Dutch Llama-3 tokens (104GB) using LoRa.
Note that this is a base model, not optimized for conversational behavior.
If this is desired for your use-case, we recommend finetuning this model on your own Dutch data or using the instruction-finetuned version of this model, [Llama-3-ChocoLlama-instruct](https://huggingface.co/ChocoLlama/Llama-3-ChocoLlama-instruct).
Use the code below to get started with the model.
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('ChocoLlama/Llama-3-ChocoLlama-base')
model = AutoModelForCausalLM.from_pretrained('ChocoLlama/Llama-3-ChocoLlama-base')
```
## Model Details
ChocoLlama is a family of open LLM's specifically adapted to Dutch, contributing to the state-of-the-art of Dutch open LLM's in their weight class.
We provide 6 variants (of which 3 base and 3 instruction-tuned models):
- **ChocoLlama-2-7B-base** ([link](https://huggingface.co/ChocoLlama/ChocoLlama-2-7B-base)): A language-adapted version of Meta's Llama-2-7b, fine-tuned on 32B Dutch Llama-2 tokens (104GB) using LoRa.
- **ChocoLlama-2-7B-instruct** ([link](https://huggingface.co/ChocoLlama/ChocoLlama-2-7B-instruct)): An instruction-tuned version of ChocoLlama-2-7B-base, fine-tuned on a collection of Dutch translations of instruction-tuning datasets, using SFT followed by DPO.
- **ChocoLlama-2-7B-tokentrans-base** ([link](https://huggingface.co/ChocoLlama/ChocoLlama-2-7B-tokentrans-base)): A language-adapted version of Meta's Llama-2-7b, using a Dutch RoBERTa-based tokenizer. The token embeddings of this model were reinitialized using the token translation algorithm proposed by [Remy et al.](https://arxiv.org/pdf/2310.03477). The model was subsequently fine-tuned on the same Dutch dataset as ChocoLlama-2-7B-base, again using LoRa.
- **ChocoLlama-2-7B-tokentrans-instruct** ([link](https://huggingface.co/ChocoLlama/ChocoLlama-2-7B-tokentrans-instruct)): An instruction-tuned version of ChocoLlama-2-7B-tokentrans-base, fine-tuned on the same dataset as ChocoLlama-2-7B-instruct, again using SFT followed by DPO.
- **Llama-3-ChocoLlama-8B-base** ([link](https://huggingface.co/ChocoLlama/Llama-3-ChocoLlama-8B-base)): A language-adapted version of Meta's Llama-8-8B, fine-tuned on the same Dutch dataset as ChocoLlama-2-7B-base, again using LoRa.
- **Llama-3-ChocoLlama-instruct** ([link](https://huggingface.co/ChocoLlama/Llama-3-ChocoLlama-8B-instruct)): An instruction-tuned version of Llama-3-ChocoLlama-8B-base, fine-tuned on the same dataset as ChocoLlama-2-7B-instruct, again using SFT followed by DPO.
For benchmark results for all models, including compared to their base models and other Dutch LLMs, we refer to our paper [here](some_url).
### Model Description
- **Developed by:** [Matthieu Meeus](https://huggingface.co/matthieumeeus97), [Anthony Rathé](https://huggingface.co/anthonyrathe)
- **Funded by:** [Vlaams Supercomputer Centrum](https://www.vscentrum.be/), through a grant of apx. 40K GPU hours (NVIDIA A100-80GB)
- **Language(s):** Dutch
- **License:** [Llama-3 Community License](https://www.llama.com/llama3/license/)
- **Finetuned from model:** [Llama-3-8b](https://huggingface.co/meta-llama/Meta-Llama-3-8B)
### Model Sources
- **Repository:** Will be released soon.
- **Paper:** Will be released soon.
## Uses
### Direct Use
Since this is a base model, we do not recommend using it for your use-cases directly. We instead recommend:
1. Fine-tuning this model to your specific use-case
2. Leveraging the instruction-tuned version of this model
### Downstream Use
Since this model is a base model, it can easily be adapted to specific use-cases that required Dutch language understanding and generation.
We expect this model to be particularly useful for use-cases in the domains which were explicitly covered in our dataset, e.g. the analysis and/or generation of Dutch job descriptions, corporate filings and legislation.
### Out-of-Scope Use
- Use-cases requiring a chat-style interface: since this is a base model, it cannot be used reliably for turn-based chat interaction. Please refer to the instruction-tuned version of this model instead.
- Use-cases requiring understanding or generation of text in languages other than Dutch: the dataset on which this model was fine-tuned does not contain data in languages other than Dutch, hence we expect significant catastrophic forgetting to have occured for English, which is the language Llama-2 was originally trained for.
## Bias, Risks, and Limitations
We have taken care to include only widely used and high-quality data in our dataset. Some of this data has been filtered by the original creators.
However we did not explicitly conduct any additional filtering of this dataset with regards to biased or otherwise harmful content.
### Recommendations
We recommend fine-tuning this model to your curated data to maximally avoid undesirable outputs.
## Training Details
### Training Data
We collect a diverse set of Dutch natural language.
1. **OSCAR**
The bulk of our data comes from the Dutch portion of [OSCAR](https://oscar-corpus.com), January 2023 version, based on Common Crawl. This dataset includes **93 GB** of text (~28.6B tokens).
2. **Open Subtitles**
We collected Dutch text from movie subtitles, focusing on unique movies either in Dutch or with Dutch subtitles. This dataset contains **5 GB** of text (~1.54B tokens) from **214k samples**.
3. **Project Gutenberg**
We downloaded **970 full Dutch books** from [Project Gutenberg](https://www.gutenberg.org) using a public scraper. The dataset includes **0.3 GB** of text (~92M tokens) and is available on [Hugging Face](https://huggingface.co/datasets/ChocoLlama/gutenberg-dutch).
4. **Wikipedia**
Using the March 2023 [Wikipedia dump](https://dumps.wikimedia.org), we included **2.5 GB** of text (~769M tokens). Despite some duplication with OSCAR, Wikipedia's high quality justifies its inclusion.
5. **Job Descriptions (TechWolf)**
A sample of **750k Dutch job descriptions** collected over five years from public websites, provided by TechWolf. This dataset contains **1.5 GB** of text (~462M tokens).
6. **Staatsblad (Bizzy)**
A sample of **80k legal filings** from [Het Belgisch Staatsblad](https://www.ejustice.just.fgov.be/cgi/welcome.pl). Documents were OCR-processed, and personal data was excluded. This dataset includes **1.4 GB** of text (~431M tokens), collected with help from Bizzy.
7. **Legislation (ML6)**
**15k documents** from Flemish legislation accessed via the [Open Data API](https://www.vlaanderen.be/vlaams-parlement/de-vlaamse-codex). This dataset contains **0.2 GB** of text (~62M tokens), collected with support from ML6.
### Training Procedure
This model was fine-tuned using low-rank (LoRa) adapatation with trainable embeddings, for a total of 1.07B trainable parameters.
#### Training Hyperparameters
- **Training regime:** bf16 non-mixed precision
- **Epochs:** 1
- **LoRa parameters:**
- R: 8
- Alpha: 32
- Trainable modules: q_proj, v_proj, k_proj, o_proj, gate_proj, up_proj, down_proj, embed_tokens, lm_head
- LoRa dropout: 0.05
- **Learning Rate:**
- Scheduler: StepLR
- Step size: 6212
- Learning rate: 0.0003
- Gamma: 0.85
- **Other parameters:**
- Minibatch size: 16
- Gradient accumulation steps: 8
- Parallelization factor: 8
- Weight decay: 0
## Evaluation
### Quantitative evaluation
We have evaluated our models on several industry-standard Dutch benchmarks, translated from their original versions. The results can be found in the table below, together with results from several other prominent Dutch models.
| Model | ARC | HellaSwag | MMLU | TruthfulQA | Avg. |
|----------------------------------------------|----------------|----------------|----------------|----------------|----------------|
| **Llama-3-ChocoLlama-instruct** | **0.48** | **0.66** | **0.49** | **0.49** | **0.53** |
| llama-3-8B-rebatch | 0.44 | 0.64 | 0.46 | 0.48 | 0.51 |
| llama-3-8B-instruct | 0.47 | 0.59 | 0.47 | 0.52 | 0.51 |
| llama-3-8B | 0.44 | 0.64 | 0.47 | 0.45 | 0.5 |
| Reynaerde-7B-Chat | 0.44 | 0.62 | 0.39 | 0.52 | 0.49 |
| **Llama-3-ChocoLlama-base** | **0.45** | **0.64** | **0.44** | **0.44** | **0.49** |
| zephyr-7b-beta | 0.43 | 0.58 | 0.43 | 0.53 | 0.49 |
| geitje-7b-ultra | 0.40 | 0.66 | 0.36 | 0.49 | 0.48 |
| **ChocoLlama-2-7B-tokentrans-instruct** | **0.45** | **0.62** | **0.34** | **0.42** | **0.46** |
| mistral-7b-v0.1 | 0.43 | 0.58 | 0.37 | 0.45 | 0.46 |
| **ChocoLlama-2-7B-tokentrans-base** | **0.42** | **0.61** | **0.32** | **0.43** | **0.45** |
| **ChocoLlama-2-7B-instruct** | **0.36** | **0.57** | **0.33** | **0.45** | **0.43 |
| **ChocoLlama-2-7B-base** | **0.35** | **0.56** | **0.31** | **0.43** | **0.41** |
| llama-2-7b-chat-hf | 0.36 | 0.49 | 0.33 | 0.44 | 0.41 |
| llama-2-7b-hf | 0.36 | 0.51 | 0.32 | 0.41 | 0.40 |
On average, Llama-3-ChocoLlama-instruct surpasses the previous state-of-the-art on these benchmarks.
### Qualitative evaluation
In our paper, we also provide an additional qualitative evaluation of all models - which we empirically find more reliable.
For details, we refer to the paper and to our benchmark [ChocoLlama-Bench](https://huggingface.co/datasets/ChocoLlama/ChocoLlama-Bench).
### Compute Infrastructure
All ChocoLlama models have been trained on the compute cluster provided by the [Flemish Supercomputer Center (VSC)](https://www.vscentrum.be/). We used 8 to 16 NVIDIA A100 GPU's with 80 GB of VRAM. |