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 | |