File size: 6,871 Bytes
baf0eff
12e189e
 
 
 
4b45c15
 
 
12e189e
 
 
 
4b45c15
 
 
12e189e
 
 
 
baf0eff
4b45c15
 
 
8a0d5a7
 
12e189e
4b45c15
5eb7542
4b45c15
12e189e
 
 
 
8a0d5a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4b45c15
8a0d5a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4b45c15
 
 
12e189e
 
 
 
 
 
4b45c15
8a0d5a7
 
 
 
 
 
 
4b45c15
 
12e189e
 
 
 
 
4b45c15
baf0eff
 
12e189e
 
 
 
4b45c15
baf0eff
4b45c15
5eb7542
4b45c15
 
 
 
 
 
 
 
 
 
12e189e
 
 
 
 
 
 
 
 
 
 
4b45c15
baf0eff
4b45c15
 
5eb7542
 
 
 
 
 
 
 
 
 
 
 
4b45c15
 
 
baf0eff
4b45c15
 
 
d5d6eb1
12e189e
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
---
license: cc-by-nc-4.0
inference: false
datasets:
- BramVanroy/alpaca-cleaned-dutch
base_model: DAMO-NLP-MT/polylm-1.7b
tags:
- generated_from_trainer
- alpaca
- Transformers
- PolyLM
- text-generation-inference
model-index:
- name: polylm_1.7b_ft_alpaca_clean_dutch
  results: []
language:
- nl
library_name: peft
pipeline_tag: text-generation
---

# polylm_1.7b_ft_alpaca_clean_dutch

## Model description

This adapter model is a fine-tuned version of [DAMO-NLP-MT/polylm-1.7b](https://huggingface.co/DAMO-NLP-MT/polylm-1.7b).
It achieves the following results on the evaluation set:
- Loss: 1.8483

Finetuning was performed on the Dutch [BramVanroy/alpaca-cleaned-dutch](https://www.huggingface.co/datasets/BramVanroy/alpaca-cleaned-dutch) dataset which contains 52K of records with instruction following-data translated from English to Dutch.

See [DAMO-NLP-MT/polylm-1.7b](https://huggingface.co/DAMO-NLP-MT/polylm-1.7b) 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 AutoModelForCausalLM, AutoTokenizer

model_name = "robinsmits/polylm_1.7b_ft_alpaca_clean_dutch"

tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast = False, legacy = False)

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

prompt = "### Instructie:\nWat zijn de drie belangrijkste softwareonderdelen die worden gebruikt bij webontwikkeling?\n\n### Antwoord:\n" 

inputs = tokenizer(prompt, return_tensors = "pt")
sample = model.generate(input_ids = inputs.input_ids.cuda(),
                        attention_mask = inputs.attention_mask.cuda(),
                        max_new_tokens = 128,
                        do_sample = True,
                        top_p = 0.85,
                        top_k = 50,
                        temperature = 0.5,
                        repetition_penalty = 1.2,
                        length_penalty = -1.0,
                        num_return_sequences = 1,
                        pad_token_id = tokenizer.eos_token_id,
                        forced_eos_token_id = tokenizer.eos_token_id)
output = tokenizer.decode(sample[0], skip_special_tokens = True)

print(output.split(prompt)[1])
```

The prompt and generated output for the above mentioned example is similar to the output shown below.

```
### Instructie:
Wat zijn de drie belangrijkste softwareonderdelen die worden gebruikt bij webontwikkeling?

### Antwoord:

De drie belangrijkste softwareonderdelen die worden gebruikt in webontwikkeling zijn HTML, CSS en Javascript.HTML is het hoofdbestand voor alle inhoud op een website.CSS is het hoofdbestand voor decoraties en scripts om te gebruiken zoals JavaScript en PHP.Javascript wordt meestal gebruikt om verschillende functies uit te voeren of het script te manipuleren.Het laatste bestand maakt het mogelijk om code te schrijven dat aan uw website gekoppeld kan worden door middel van enkele woorden. Daarnaast kunnen er ook andere bestanden nodig zijn als gevolg van gebruik van meerdere servers.Een voorbeeld hiervan zou zijn wanneer u bijvoorbeeld een blog-website
```

For more extensive usage and a lot of generated samples (both good and bad samples) see the following [Inference Notebook](https://github.com/RobinSmits/Dutch-LLMs/blob/main/PolyLM_1_7B_Alpaca_Clean_Dutch_Inference.ipynb)

## Intended uses & limitations

The PolyLM-1.7B model was trained on 18 languages. The primary focus was to create a multi-lingual Open LLM.
Dutch was one of those 18 languages. For training the model a diverse combination of multi-lingual datasets was used. 

The generated output and performance of this model for the Dutch language is very likely not always comparable to the various Open-Llama models that have been finetuned on English Alpaca datasets.

The primary intention of this finetuned model is to explore and research the use of the Dutch language in combination with an Open LLM model.

## Bias, Risks, and Limitations

The information below is copied from the base model's [official model card](https://arxiv.org/pdf/2307.06018.pdf):
This applies also to the finetuned model.

> Our contributions are fully methodological: adding the support of multilingualism to LLM during training and SFT phases. It is unavoidable that PolyLM might exhibit several common deficiencies of language models, e.g. hallucination and toxicity. PolyLM should not be used directly in any application, without a prior assessment of safety and fairness concerns specific to the application.

## Training and evaluation data

This model was trained on the [BramVanroy/alpaca-cleaned-dutch](https://www.huggingface.co/datasets/BramVanroy/alpaca-cleaned-dutch) dataset.

The dataset is the Dutch translation of the English Alpaca Cleaned instruction dataset.

Based on the dataset license only Non-Commercial use is allowed. Commercial use is strictly forbidden.

## Training procedure

This model was finetuned with a QLoRA setup on a Google Colab A100 GPU in about 1.5 hours.

The notebook used for training can be found here: [Training Notebook](https://github.com/RobinSmits/Dutch-LLMs/blob/main/PolyLM_1_7B_Alpaca_Clean_Dutch_Qlora.ipynb)

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 8
- total_train_batch_size: 64
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 64
- num_epochs: 2

The following bitsandbytes quantization config was used during training:
- load_in_8bit: False
- load_in_4bit: True
- llm_int8_threshold: 6.0
- llm_int8_skip_modules: None
- llm_int8_enable_fp32_cpu_offload: False
- llm_int8_has_fp16_weight: False
- bnb_4bit_quant_type: nf4
- bnb_4bit_use_double_quant: True
- bnb_4bit_compute_dtype: bfloat16

### Training results

| Training Loss | Epoch | Step | Validation Loss |
|:-------------:|:-----:|:----:|:---------------:|
| 2.1248        | 0.16  | 128  | 2.1129          |
| 2.0512        | 0.33  | 256  | 2.0347          |
| 1.9983        | 0.49  | 384  | 1.9948          |
| 1.9557        | 0.66  | 512  | 1.9655          |
| 1.9583        | 0.82  | 640  | 1.9386          |
| 1.916         | 0.99  | 768  | 1.9177          |
| 1.8671        | 1.15  | 896  | 1.9019          |
| 1.8626        | 1.32  | 1024 | 1.8885          |
| 1.8321        | 1.48  | 1152 | 1.8762          |
| 1.8596        | 1.65  | 1280 | 1.8631          |
| 1.843         | 1.81  | 1408 | 1.8539          |
| 1.8333        | 1.98  | 1536 | 1.8483          |


### Framework versions

- Transformers 4.31.0
- Pytorch 2.0.1+cu118
- Datasets 2.13.1
- Tokenizers 0.13.3
- PEFT 0.4.0