|
--- |
|
license: apache-2.0 |
|
base_model: Rijgersberg/GEITje-7B |
|
tags: |
|
- generated_from_trainer |
|
- GEITje |
|
- conversational |
|
model-index: |
|
- name: GEITje-7B-chat-v2 |
|
results: [] |
|
datasets: |
|
- Rijgersberg/no_robots_nl |
|
- Rijgersberg/ultrachat_10k_nl |
|
- BramVanroy/dutch_chat_datasets |
|
language: |
|
- nl |
|
pipeline_tag: text-generation |
|
--- |
|
# GEITje-7B-chat-v2 |
|
|
|
**🤖️ Try the chat model in [🤗 Hugging Face Spaces](https://huggingface.co/spaces/Rijgersberg/GEITje-7B-chat)!** |
|
|
|
# GEITje-7B |
|
|
|
GEITje is a large open Dutch language model with 7 billion parameters, based on Mistral 7B. |
|
It has been further trained on 10 billion tokens of Dutch text. |
|
This has improved its Dutch language skills and increased its knowledge of Dutch topics. |
|
|
|
|
|
## Model description |
|
|
|
### _Mistral_ – Base Model |
|
GEITje is based on [Mistral 7B](https://mistral.ai/news/announcing-mistral-7b/). |
|
It's a large open language model with 7 billion parameters, |
|
trained by [Mistral AI](https://mistral.ai). |
|
According to Mistral AI, the 7B model performs better than [Llama 2](https://ai.meta.com/llama/) 13B on all (English-language) benchmarks they tested it on. |
|
Mistral 7B has been released under the Apache 2.0 open source license. |
|
|
|
|
|
### _GEITje_ – Trained Further on Dutch Texts |
|
GEITje was created by further training Mistral 7B on no less than 10 billion tokens of Dutch text from the [Dutch Gigacorpus](http://gigacorpus.nl) and the [MADLAD-400](https://huggingface.co/datasets/allenai/MADLAD-400) web crawling corpus. |
|
It is a so-called _full-parameter finetune_: |
|
performed on all parameters. |
|
It is not a [PEFT](https://huggingface.co/blog/peft) or [LoRA](https://huggingface.co/docs/peft/conceptual_guides/lora) finetune. |
|
Like Mistral, GEITje has a _context length_ of 8,192 tokens. |
|
|
|
### _GEITje-chat_ – Finetuned for Dialogues |
|
As a demonstration of GEITje's capabilities for chat applications, two initial chat variants of GEITje have also been finetuned: GEITje-chat and GEITje-chat-v2. |
|
They can follow instructions, answer questions, and hold dialogues on a variety of topics. |
|
|
|
## More info |
|
Read more about GEITje-chat in the [📄 README](https://github.com/Rijgersberg/GEITje/blob/main/README-en.md) on GitHub. |
|
|
|
## Checkpoints |
|
An intermediate checkpoint is available in the `checkpoints` branch. |
|
|
|
## Training procedure |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 5e-06 |
|
- train_batch_size: 2 |
|
- eval_batch_size: 8 |
|
- seed: 42 |
|
- gradient_accumulation_steps: 8 |
|
- total_train_batch_size: 16 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: cosine |
|
- lr_scheduler_warmup_ratio: 0.1 |
|
- num_epochs: 1 |
|
|
|
### Training results |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | |
|
|:-------------:|:-----:|:-----:|:---------------:| |
|
| 0.7832 | 0.05 | 609 | 0.8844 | |
|
| 0.6904 | 0.1 | 1218 | 0.8698 | |
|
| 0.8195 | 0.15 | 1827 | 0.8583 | |
|
| 0.7463 | 0.2 | 2436 | 0.8475 | |
|
| 0.6739 | 0.25 | 3045 | 0.8395 | |
|
| 0.7604 | 0.3 | 3654 | 0.8332 | |
|
| 0.8024 | 0.35 | 4263 | 0.8261 | |
|
| 0.6881 | 0.4 | 4872 | 0.8203 | |
|
| 0.6466 | 0.45 | 5481 | 0.8167 | |
|
| 0.7042 | 0.5 | 6090 | 0.8121 | |
|
| 0.702 | 0.55 | 6699 | 0.8081 | |
|
| 0.7255 | 0.6 | 7308 | 0.8054 | |
|
| 0.7558 | 0.65 | 7917 | 0.8036 | |
|
| 0.7587 | 0.7 | 8526 | 0.8022 | |
|
| 0.9217 | 0.75 | 9135 | 0.8016 | |
|
| 0.6938 | 0.8 | 9744 | 0.8011 | |
|
| 0.6962 | 0.85 | 10353 | 0.8011 | |
|
| 0.664 | 0.9 | 10962 | 0.8011 | |
|
| 0.6544 | 0.95 | 11571 | 0.8011 | |
|
| 0.6782 | 1.0 | 12180 | 0.8011 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.36.0.dev0 |
|
- Pytorch 2.1.1+cu121 |
|
- Datasets 2.15.0 |
|
- Tokenizers 0.15.0 |
