BramVanroy/fietje-2-chat

Fietje 2 Chat

An open and efficient LLM for Dutch

👱‍♀️ Base version - 🤖 Instruct version - 💬 Chat version (this one) - 🚀 GGUF of Chat

Chat with Fietje here!

This is the chat version of Fietje, a DPO-tuned (aligned) continuation on the instruct version. Fietje is an adapated version of microsoft/phi-2, tailored to Dutch text generation by training on 28B tokens. It is small and efficient with a size of 2.7 billion parameters while performing almost on par with more powerful Dutch LLMs of twice its size like GEITje 7B Ultra.

A thorough description of the creation and evaluation of Fietje as well as usage examples are available in this Github repository.

Citation

If you use Fietje or the CulturaX + Wikipedia filtered subset in your work, please cite to the following paper:

@misc{vanroy2024fietjeopenefficientllm,
      title={Fietje: An open, efficient LLM for Dutch}, 
      author={Bram Vanroy},
      year={2024},
      eprint={2412.15450},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.15450}, 
}

Intended uses & limitations

The same limitations as phi-2, and LLMs in general, apply here. LLMs hallucinate, make mistakes, and should not be trusted. Use at your own risk!

Training and evaluation data

Fietje 2 Chat was finetuned from the instruct model on the following datasets. Number of training samples per dataset given in brackets, totalling 18,653 samples.

• BramVanroy/ultra_feedback_dutch_cleaned subset dpo_hq: a cleaned version of BramVanroy/ultra_feedback_dutch (9186)
• BramVanroy/orca_dpo_pairs_dutch_cleaned subset dpo_all: a cleaned version of BramVanroy/orca_dpo_pairs_dutch (9467)

A lot of different learning rates, beta, en batch sizes were investigated in search of a converging combination. You can find them all in the W&B runs.

Training procedure

I am thankful to the Flemish Supercomputer Center (VSC) for providing the computational power to accomplish this project. Accounting for waiting for jobs, training a single run took around nine hours on one A100 80GB.

Training was done with the wonderful alignment-handbook, using DeepSpeed as a back-end. Exact training recipes and SLURM script are given in the Github repository.

Training hyperparameters

The following hyperparameters were used during training:

• beta: 0.2
• learning_rate: 2e-06
• train_batch_size: 8
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-07
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.2515	1.0	1166	0.2842	-1.1549	-3.6363	0.8867	2.4815	-657.6813	-451.3364	-1.2868	-1.3528

Framework versions

• Transformers 4.39.1
• Pytorch 2.1.2+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2