File size: 11,972 Bytes
8cb54a2
 
4b0c0be
 
8cb54a2
 
a7191ef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8cb54a2
 
a7191ef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
---
library_name: transformers
license: apache-2.0
basemodel: Qwen/Qwen1.5-7B
---

## Unsloth x Qwen2
[Unsloth](https://github.com/unslothai/unsloth) can speed up training LLM and reduce memory usage, but currently it only supports Llama3, Mistral, Gemma, ORPR, Phi-3 and TinyLlama.
We can't train Qwen2 with Unsloth, even though Qwen2 is popular in community. 

It's exciting that we succeed to make Unsloth support Qwen2, it can speed up training and reduce much memory usage. 
If you want to train Qwen2 with Unsloth, you can use [our repo](https://github.com/yangjianxin1/unsloth) rather than the official one. And we will commit our code to the [official repo](https://github.com/unslothai/unsloth).

Install our Unsloth:
```bash
pip install git+https://github.com/yangjianxin1/unsloth.git
```

[Firefly](https://github.com/yangjianxin1/Firefly) already supports training Qwen2 with Unsloth, and the subsequent models are trained with Firefly, you can try it.


## Model Card for Firefly-Qwen1.5-Unsloth
[firefly-qwen1.5-en-7b-unsloth](https://huggingface.co/YeungNLP/firefly-qwen1.5-en-7b-unsloth) and [firefly-qwen1.5-en-7b-dpo-v0.1-unloth](https://huggingface.co/YeungNLP/firefly-qwen1.5-en-7b-dpo-v0.1-unsloth) are trained based on [Qwen1.5-7B](https://huggingface.co/Qwen/Qwen1.5-7B) to act as a helpful and harmless AI assistant. 
We use [Firefly](https://github.com/yangjianxin1/Firefly) to train our models on **a single V100 GPU** with QLoRA and [Unsloth](https://github.com/yangjianxin1/unsloth).
firefly-qwen1.5-en-7b-unsloth is fine-tuned based on Qwen1.5-7B with English instruction data, and firefly-qwen1.5-en-7b-dpo-v0.1-unsloth is trained with [Direct Preference Optimization (DPO)](https://arxiv.org/abs/2305.18290) based on firefly-qwen1.5-en-7b-unsloth.

Our models outperform official [Qwen1.5-7B-Chat](https://huggingface.co/Qwen/Qwen1.5-7B-Chat), [Gemma-7B-it](https://huggingface.co/google/gemma-7b-it), [Zephyr-7B-Beta](https://huggingface.co/HuggingFaceH4/zephyr-7b-beta) on [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard).

Although our models are trained with English data, you can also try to chat with models in Chinese because Qwen1.5 is also good at Chinese. But we have not evaluated
the performance in Chinese yet.

We advise you to install transformers>=4.37.0.

## Performance
We have evaluated the training gain of Qwen1.5-7B, we use QLoRA and Unsloth to train model for 20 steps on a single V100. The result can be listed as follows.
**Unsloth can reduce GPU memory by 39.13% and training time by 32.12%, and the training speed can increase by 47.32%.**

| max_seq_length | per_device_train_batch_size | gradient_accumulation_steps | use_unsloth | rank | GPU                     | Time              |
|----------------|----------------------------|-----------------------------|-------------|------|-------------------------|-------------------|
| 1024           | 1                          | 16                          | false       | 8    | 13.72GB                 | 448s              |
| 1024           | 1                          | 16                          | true        | 8    | **8.43GB**(**-38.56%**) | 308s(**-31.25%**) |
| 1024           | 1                          | 16                          | false       | 64   | 16.01GB                 | 452s              |
| 1024           | 1                          | 16                          | true        | 64   | 11.07GB(**-30.86%**)    | 311s(**-31.19%**) |
| 2048           | 1                          | 16                          | false       | 64   | 18.55GB                 | 840s              |
| 2048           | 1                          | 16                          | true        | 64   | 12.99GB(**-29.97%**)    | 596s(**-29.05%**) |
| 1024           | 4                          | 4                           | false       | 64   | 24.70GB                 | 357s              |
| 1024           | 4                          | 4                           | true        | 64   | 14.36GB(**-41.86%**)    | 253s(**-29.13%**) |
| 2048           | 4                          | 4                           | false       | 64   | 32.51GB                 | 741s              |
| 2048           | 4                          | 4                           | true        | 64   | 19.79GB(**-39.13%**)    | 503s(**-32.12%**) |


We evaluate our sft and dpo models on [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard), they achieve good performance.

| Model                                      | Average | ARC    | HellaSwag | MMLU  | TruthfulQA | Winogrande | GSM8K  |
|--------------------------------------------|---------|--------|-----------|-------|------------|------------|--------|
| firefly-gemma-7b                           | 62.93   | 	62.12 | 79.77     | 61.57 | 49.41      | 75.45      | 49.28  |
| **firefly-qwen1.5-en-7b-dpo-v0.1-unsloth** | 62.65   | 56.14  | 75.5      | 60.87 | 58.09      | 70.72      | 54.59  |
| zephyr-7b-beta                             | 61.95   | 62.03  | 84.36     | 61.07 | 	57.45     | 77.74      | 	29.04 |
| **firefly-qwen1.5-en-7b-unsloth**                 | 61.81   | 54.27  | 	76.22    | 61.55 | 50.62      | 70.48      | 57.7   |
| vicuna-13b-v1.5                            | 55.41   | 57.08  | 	81.24    | 56.67 | 51.51      | 	74.66     | 11.3   |
| Xwin-LM-13B-V0.1                           | 55.29   | 	62.54 | 82.8      | 56.53 | 45.96      | 74.27      | 9.63   |
| Qwen1.5-7B-Chat                            | 55.15   | 	55.89 | 78.56     | 61.65 | 53.54      | 	67.72     | 13.57  |
| gemma-7b-it                                | 53.56   | 51.45  | 71.96     | 53.52 | 47.29      | 	67.96     | 	29.19 |



## Usage
The chat templates of our chat models are the same as Official Qwen1.5-7B-Chat:
```text
<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
hello, who are you?<|im_end|>
<|im_start|>assistant
I am a AI program developed by Firefly<|im_end|>
```

You can use script to inference in [Firefly](https://github.com/yangjianxin1/Firefly/blob/master/script/chat/chat.py).

You can also use the following code:
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name_or_path = "YeungNLP/firefly-qwen1.5-en-7b-unsloth"
model = AutoModelForCausalLM.from_pretrained(
    model_name_or_path,
    trust_remote_code=True,
    low_cpu_mem_usage=True,
    torch_dtype=torch.float16,
    device_map='auto',
)
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)

prompt = "Compose an engaging travel blog post about a recent trip to Hawaii, highlighting cultural experiences and must-see attractions. "
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to('cuda')

generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=1500,
    top_p = 0.9,
    temperature = 0.35,
    repetition_penalty = 1.0,
    eos_token_id=tokenizer.encode('<|im_end|>', add_special_tokens=False)
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)
```

## Training Details
Both in SFT and DPO stages, **We only use a single V100 GPU** with QLoRA and Unsloth, and we use [Firefly](https://github.com/yangjianxin1/Firefly) to train our models.

### Training Setting
The following hyperparameters are used during SFT:
- num_epochs: 1
- learning_rate: 2e-4
- total_train_batch_size: 32
- max_seq_length: 2048
- optimizer: paged_adamw_32bit
- lr_scheduler_type: constant_with_warmup
- warmup_steps: 600
- lora_rank: 64
- lora_alpha: 16
- lora_dropout: 0.05
- gradient_checkpointing: true
- fp16: true

The following hyperparameters were used during DPO:
- num_epochs: 1
- learning_rate: 2e-4
- total_train_batch_size: 32
- max_seq_length: 2048
- max_prompt_length: 500
- optimizer: paged_adamw_32bit
- lr_scheduler_type: constant_with_warmup
- warmup_steps: 100
- lora_rank: 64
- lora_alpha: 16
- lora_dropout: 0.05
- gradient_checkpointing: true
- fp16: true


### Training metrics

The table below shows the full set of DPO training metrics:

| Epoch | Step | Loss   | Rewards/accuracies | Rewards/margins | Rewards/chosen | Rewards/rejected | Logits/chosen | Logits/rejected | Logps/chosen | Logps/rejected | 
|-------|------|--------|--------------------|-----------------|----------------|------------------|---------------|-----------------|--------------|----------------|
| 0.05  | 100  | 0.6128 | 0.6572             | 0.3914          | -0.0622        | -0.4537          | 1.107         | 1.1104          | -283.7632    | -264.5925      |
| 0.1   | 200  | 0.6066 | 0.6913             | 0.662           | -0.3589        | -1.0209          | 0.9433        | 0.9431          | -279.0002    | -268.6432      |
| 0.16  | 300  | 0.5803 | 0.7069             | 0.876           | -0.3849        | -1.2609          | 0.8411        | 0.8537          | -289.9482    | -274.3425      |
| 0.21  | 400  | 0.5624 | 0.7169             | 0.9575          | -0.2447        | -1.2022          | 0.7615        | 0.7497          | -293.8072    | -274.4167      |
| 0.26  | 500  | 0.5863 | 0.7                | 0.8908          | -0.5283        | -1.4191          | 0.537         | 0.5085          | -284.3388    | -267.9294      |
| 0.31  | 600  | 0.5612 | 0.7166             | 1.0791          | -0.592         | -1.6711          | 0.7121        | 0.7219          | -293.2425    | -278.5992      |
| 0.37  | 700  | 0.5741 | 0.7234             | 1.0742          | -0.8469        | -1.9211          | 0.6002        | 0.5769          | -300.8099    | -285.9137      |
| 0.42  | 800  | 0.582  | 0.7141             | 1.0414          | -1.1658        | -2.2072          | 0.7191        | 0.5934          | -300.458     | -286.1         |
| 0.47  | 900  | 0.5694 | 0.7178             | 1.2055          | -1.7372        | -2.9426          | 0.4226        | 0.316           | -305.5303    | -290.7548      |
| 0.52  | 1000 | 0.5827 | 0.7134             | 1.1063          | -1.354         | -2.4603          | 0.535         | 0.4022          | -302.7598    | -286.636       |
| 0.58  | 1100 | 0.5553 | 0.7306             | 1.3631          | -1.5861        | -2.9492          | 0.7636        | 0.6559          | -312.9375    | -290.3474      |
| 0.63  | 1200 | 0.5633 | 0.7341             | 1.2689          | -1.7187        | -2.9876          | 0.6555        | 0.5894          | -315.0179    | -298.2406      |
| 0.68  | 1300 | 0.5705 | 0.7284             | 1.3501          | -1.7762        | -3.1263          | 0.7419        | 0.6874          | -310.9056    | -294.2934      |
| 0.73  | 1400 | 0.5458 | 0.7347             | 1.4555          | -2.2377        | -3.6932          | 0.7279        | 0.6564          | -309.141     | -299.1613      |
| 0.79  | 1500 | 0.5797 | 0.7222             | 1.2937          | -2.4483        | -3.742           | 0.8444        | 0.771           | -321.578     | -298.111       |
| 0.84  | 1600 | 0.5572 | 0.7319             | 1.4824          | -2.9344        | -4.4168          | 0.9202        | 0.8605          | -323.4034    | -307.0114      |
| 0.89  | 1700 | 0.5518 | 0.7281             | 1.4263          | -2.7301        | -4.1564          | 0.9257        | 0.8785          | -313.694     | -298.1267      |
| 0.94  | 1800 | 0.5572 | 0.7272             | 1.5121          | -2.9505        | -4.4627          | 0.7899        | 0.7503          | -314.1552    | -305.9873      |
| 0.99  | 1900 | 0.5763 | 0.7241             | 1.4982          | -2.7064        | -4.2047          | 0.7841        | 0.7023          | -310.6677    | -299.5064      |