language:
- nl
license: apache-2.0
tags:
- generated_from_trainer
- GEITje
datasets:
- Rijgersberg/GEITje-pretrain-10b
base_model: mistralai/Mistral-7B-v0.1
model-index:
- name: GEITje-v1-7B
results: []
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. It's a large open language model with 7 billion parameters, trained by Mistral AI. According to Mistral AI, the 7B model performs better than Llama 2 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 and the MADLAD-400 web crawling corpus. It is a so-called full-parameter finetune: performed on all parameters. It is not a PEFT or LoRA finetune. Like Mistral, GEITje has a context length of 8,192 tokens.
More info
Read more about GEITje in the 📄 README on GitHub.
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 2
- eval_batch_size: 2
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- gradient_accumulation_steps: 8
- total_train_batch_size: 128
- total_eval_batch_size: 16
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 953
- training_steps: 9536
Training results
| Training Loss | Epoch | Step | Validation Loss |
|---|---|---|---|
| 1.6995 | 0.02 | 199 | 1.7673 |
| 1.6949 | 0.04 | 398 | 1.6880 |
| 1.6377 | 0.06 | 597 | 1.6429 |
| 1.6011 | 0.08 | 796 | 1.6384 |
| 1.5196 | 0.1 | 995 | 1.6060 |
| 1.5158 | 0.13 | 1194 | 1.5832 |
| 1.5181 | 0.15 | 1393 | 1.5541 |
| 1.4931 | 0.17 | 1592 | 1.5493 |
| 1.4972 | 0.19 | 1791 | 1.5407 |
| 1.5349 | 0.21 | 1990 | 1.5305 |
| 1.5025 | 0.23 | 2189 | 1.5263 |
| 1.396 | 0.25 | 2388 | 1.5140 |
| 1.4353 | 0.27 | 2587 | 1.5104 |
| 1.4307 | 0.29 | 2786 | 1.5003 |
| 1.3974 | 0.31 | 2985 | 1.4849 |
| 1.404 | 0.33 | 3184 | 1.4771 |
| 1.4299 | 0.35 | 3383 | 1.4825 |
| 1.4342 | 0.38 | 3582 | 1.4705 |
| 1.4341 | 0.4 | 3781 | 1.4643 |
| 1.4535 | 0.42 | 3980 | 1.4580 |
| 1.4799 | 0.44 | 4179 | 1.4521 |
| 1.35 | 0.46 | 4378 | 1.4478 |
| 1.4586 | 0.48 | 4577 | 1.4425 |
| 1.3685 | 0.5 | 4776 | 1.4368 |
| 1.4572 | 0.52 | 4975 | 1.4313 |
| 1.3293 | 0.54 | 5174 | 1.4265 |
| 1.403 | 0.56 | 5373 | 1.4241 |
| 1.3057 | 0.58 | 5572 | 1.4188 |
| 1.244 | 0.61 | 5771 | 1.4178 |
| 1.3224 | 0.63 | 5970 | 1.4110 |
| 1.3238 | 0.65 | 6169 | 1.4083 |
| 1.3262 | 0.67 | 6368 | 1.4050 |
| 1.3237 | 0.69 | 6567 | 1.4027 |
| 1.0453 | 0.71 | 6766 | 1.4005 |
| 1.3136 | 0.73 | 6965 | 1.3992 |
| 1.3137 | 0.75 | 7164 | 1.3975 |
| 1.1587 | 0.77 | 7363 | 1.3964 |
| 1.316 | 0.79 | 7562 | 1.3957 |
| 1.2738 | 0.81 | 7761 | 1.3951 |
| 1.308 | 0.83 | 7960 | 1.3949 |
| 1.4049 | 0.86 | 8159 | 1.3946 |
| 1.3324 | 0.88 | 8358 | 1.3944 |
| 1.3446 | 0.9 | 8557 | 1.3944 |
| 1.2489 | 0.92 | 8756 | 1.3943 |
| 1.2687 | 0.94 | 8955 | 1.3943 |
| 1.3293 | 0.96 | 9154 | 1.3943 |
| 1.3045 | 0.98 | 9353 | 1.3943 |
Framework versions
- Transformers 4.36.0.dev0
- Pytorch 2.1.1+cu121
- Datasets 2.15.0
- Tokenizers 0.15.0
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 50.53 |
| AI2 Reasoning Challenge (25-Shot) | 44.80 |
| HellaSwag (10-Shot) | 75.31 |
| MMLU (5-Shot) | 50.10 |
| TruthfulQA (0-shot) | 40.45 |
| Winogrande (5-shot) | 72.38 |
| GSM8k (5-shot) | 20.17 |