This model is a fine-tuned version of ReBatch/Reynaerde-7B-Instruct on ReBatch/ultrafeedback_nl. This is a combination of a translation of the HuggingFaceH4/ultrafeedback_binarized dataset and the HQ samples from BramVanroy's translation.
Model description
This model is a Dutch chat model, originally developed from Mistral 7B v0.3 Instruct and further fine-tuned with QLoRA. It was first fine-tuned with SFT on a chat dataset and then with DPO on a feedback chat dataset.
Intended uses & limitations
This model could still generate wrong, misleading, and potentially even offensive content. Use at your own risk. Use with Mistral's chat template (can be found in the tokenizer).
Training procedure
This model was trained with QLoRa in bfloat16 with Flash Attention 2 on one A100 PCIe, using the DPO script from the alignment handbook on RunPod.
Evaluation results
The model was evaluated using scandeval. There are improvements in 4 out of 7 benchmarks compared to the Mistral-7B-v0.3-Instruct model on which it is based.
| Model | conll_nl | dutch_social | scala_nl | squad_nl | wiki_lingua_nl | mmlu_nl | hellaswag_nl |
|---|---|---|---|---|---|---|---|
| Reynaerde-7B-Chat | 56.40 / 38.13 | 10.83 / 27.67 | 20.02 / 55.40 | 53.56 / 65.29 | 68.13 / 20.85 | 32.50 / 49.10 | 31.36 / 47.79 |
| Mistral-7B-v0.3 | 57.08 / 42.65 | 14.05 / 39.13 | 8.08 / 43.07 | 45.57 / 55.20 | 62.28 / 16.46 | 20.39 / 40.03 | 13.28 / 34.13 |
| Mistral-7B-v0.3-Instruct | 60.76 / 45.39 | 13.20 / 34.26 | 23.23 / 59.26 | 48.94 / 60.13 | 66.09 / 18.02 | 24.95 / 43.67 | 24.86 / 43.57 |
Naming
This model is named after the Middle Dutch epic poem 'Van den vos Reynaerde'. Dating from around 1260, this epic by Flemish author Willem die Madocke maecte is often called 'the pinnacle of Gothic literature in the Netherlands'. The poem tells a version of the Reynard the Fox story, popular in Western Europe during the late Middle Ages
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5e-06
- train_batch_size: 3
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- gradient_accumulation_steps: 2
- total_train_batch_size: 6
- 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
Framework versions
- PEFT 0.11.1
- Transformers 4.41.2
- Pytorch 2.2.0+cu121
- Datasets 2.19.1
- Tokenizers 0.19.1
Model Developer
The Mistral-7B-v0.3-Instruct model, on which this model is based, was created by Mistral AI. The finetuning was done by Julien Van den Avenne.