metadata

language:
  - nl
license: cc-by-nc-4.0
library_name: peft
tags:
  - trl
  - dpo
  - conversational
  - generated_from_trainer
  - qwen2
base_model: Qwen/Qwen1.5-7B-Chat
datasets:
  - BramVanroy/ultra_feedback_dutch_cleaned
pipeline_tag: text-generation
inference: false
model-index:
  - name: Qwen1.5-7B-Dutch-Chat-Dpo
    results: []

Qwen1.5-7B-Dutch-Chat-Dpo

Model description

This DPO aligned model is an adapter model based on robinsmits/Qwen1.5-7B-Dutch-Chat-Sft.

DPO Finetuning was performed on the Dutch BramVanroy/ultra_feedback_dutch_cleaned dataset.

See Qwen/Qwen1.5-7B-Chat for all information about the base model.

Model usage

A basic example of how to use the finetuned model.

import torch
from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

device = 'cuda'
model_name = 'robinsmits/Qwen1.5-7B-Dutch-Chat-Dpo'

model = AutoPeftModelForCausalLM.from_pretrained(model_name, 
                                                 device_map = "auto", 
                                                 load_in_4bit = True,
                                                 torch_dtype = torch.bfloat16)

tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [{"role": "user", "content": "Hoi hoe gaat het ermee? Wat kun je me vertellen over appels?"}]

encoded_ids = tokenizer.apply_chat_template(messages, 
                                            add_generation_prompt = True,
                                            return_tensors = "pt")

generated_ids = model.generate(input_ids = encoded_ids.to(device), 
                               max_new_tokens = 256,
                               do_sample = True)
decoded = tokenizer.batch_decode(generated_ids)
print(decoded[0])

Below the chat template with the generated output.

<|im_start|>system
Je bent een behulpzame AI assistent<|im_end|>
<|im_start|>user
Hoi hoe gaat het ermee? Wat kun je me vertellen over appels?<|im_end|>
<|im_start|>assistant
Hallo! Appels zijn zoet, knapperig en hebben een mooie smaak. Ze zijn groen met roze tinten en er zijn verschillende soorten zoals de Granny Smith, Red Delicious en Gala. Er wordt gezegd dat appels goed zijn voor je gezondheid omdat ze veel vezels en vitamines bevatten. Ook kunnen ze lekker worden gegeten zonder te bakken of te koken. Er zijn ook veel verschillende dingen die je met appels kunt doen zoals het maken van appeltaart of het drinken van appelcider.<|im_end|>

Intended uses & limitations

As with all LLM's this model can also experience bias and hallucinations. Regardless of how you use this model always perform the necessary testing and validation.

The used dataset does not allow commercial usage.

Training and evaluation data

The training notebook is available at the following link: Qwen1_5_7B_Dutch_Chat_DPO

Training was performed with Google Colab PRO on a A100 - 40GB and lasted around 4 hours.

It achieves the following results on the evaluation set:

Loss: 0.2610
Rewards/chosen: -0.7248
Rewards/rejected: -2.6224
Rewards/accuracies: 0.9170
Rewards/margins: 1.8976
Logps/rejected: -877.8102
Logps/chosen: -783.4282
Logits/rejected: -0.8110
Logits/chosen: -0.7528

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 1e-05
train_batch_size: 1
eval_batch_size: 2
seed: 42
gradient_accumulation_steps: 32
total_train_batch_size: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_ratio: 0.05
num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.5503	0.1	30	0.4684	-0.0439	-0.6295	0.8919	0.5856	-837.9513	-769.8103	-0.9335	-0.8894
0.4178	0.2	60	0.3568	-0.3713	-1.4769	0.9015	1.1056	-854.9000	-776.3594	-0.8768	-0.8276
0.3264	0.29	90	0.3143	-0.4893	-1.8730	0.9151	1.3837	-862.8228	-778.7191	-0.8428	-0.7929
0.2999	0.39	120	0.2885	-0.6832	-2.3118	0.9151	1.6286	-871.5981	-782.5971	-0.8260	-0.7730
0.3454	0.49	150	0.2749	-0.7239	-2.4904	0.9189	1.7664	-875.1693	-783.4113	-0.8235	-0.7678
0.3354	0.59	180	0.2685	-0.6775	-2.4859	0.9170	1.8084	-875.0795	-782.4824	-0.8130	-0.7574
0.2848	0.68	210	0.2652	-0.7157	-2.5692	0.9131	1.8535	-876.7465	-783.2466	-0.8157	-0.7586
0.3437	0.78	240	0.2621	-0.7233	-2.6091	0.9151	1.8857	-877.5430	-783.3994	-0.8138	-0.7561
0.2655	0.88	270	0.2611	-0.7183	-2.6154	0.9151	1.8971	-877.6708	-783.2995	-0.8106	-0.7524
0.3442	0.98	300	0.2610	-0.7248	-2.6224	0.9170	1.8976	-877.8102	-783.4282	-0.8110	-0.7528

Framework versions

PEFT 0.9.0
Transformers 4.38.2
Pytorch 2.2.1+cu121
Datasets 2.17.1
Tokenizers 0.15.2

Citation

Thanks to the creators of Qwen1.5 for there great work!

@article{qwen,
  title={Qwen Technical Report},
  author={Jinze Bai and Shuai Bai and Yunfei Chu and Zeyu Cui and Kai Dang and Xiaodong Deng and Yang Fan and Wenbin Ge and Yu Han and Fei Huang and Binyuan Hui and Luo Ji and Mei Li and Junyang Lin and Runji Lin and Dayiheng Liu and Gao Liu and Chengqiang Lu and Keming Lu and Jianxin Ma and Rui Men and Xingzhang Ren and Xuancheng Ren and Chuanqi Tan and Sinan Tan and Jianhong Tu and Peng Wang and Shijie Wang and Wei Wang and Shengguang Wu and Benfeng Xu and Jin Xu and An Yang and Hao Yang and Jian Yang and Shusheng Yang and Yang Yao and Bowen Yu and Hongyi Yuan and Zheng Yuan and Jianwei Zhang and Xingxuan Zhang and Yichang Zhang and Zhenru Zhang and Chang Zhou and Jingren Zhou and Xiaohuan Zhou and Tianhang Zhu},
  journal={arXiv preprint arXiv:2309.16609},
  year={2023}
}

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	53.94
AI2 Reasoning Challenge (25-Shot)	50.77
HellaSwag (10-Shot)	74.24
MMLU (5-Shot)	60.70
TruthfulQA (0-shot)	42.37
Winogrande (5-shot)	68.11
GSM8k (5-shot)	27.45