Hugging Face's logo

RichardErkhov
/
YeungNLP_-_firefly-qwen1.5-en-7b-4bits

Safetensors
qwen2
4-bit precision
bitsandbytes
Model card Files Community
RichardErkhov commited on
Commit
e109795
•
1 Parent(s): 2bfedf3

uploaded readme

Browse files
Files changed (1) hide show
  1. README.md +181 -0
README.md ADDED
@@ -0,0 +1,181 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Quantization made by Richard Erkhov.
2
+
3
+ [Github](https://github.com/RichardErkhov)
4
+
5
+ [Discord](https://discord.gg/pvy7H8DZMG)
6
+
7
+ [Request more models](https://github.com/RichardErkhov/quant_request)
8
+
9
+
10
+ firefly-qwen1.5-en-7b - bnb 4bits
11
+ - Model creator: https://huggingface.co/YeungNLP/
12
+ - Original model: https://huggingface.co/YeungNLP/firefly-qwen1.5-en-7b/
13
+
14
+
15
+
16
+
17
+ Original model description:
18
+ ---
19
+ library_name: transformers
20
+ license: apache-2.0
21
+ basemodel: Qwen/Qwen1.5-7B
22
+ ---
23
+
24
+ ## Model Card for Firefly-Qwen1.5
25
+
26
+ [firefly-qwen1.5-en-7b](https://huggingface.co/YeungNLP/firefly-qwen1.5-en-7b) and [firefly-qwen1.5-en-7b-dpo-v0.1](https://huggingface.co/YeungNLP/firefly-qwen1.5-en-7b-dpo-v0.1) are trained based on [Qwen1.5-7B](https://huggingface.co/Qwen/Qwen1.5-7B) to act as a helpful and harmless AI assistant.
27
+ We use [Firefly](https://github.com/yangjianxin1/Firefly) to train our models on **a single V100 GPU** with QLoRA.
28
+ firefly-qwen1.5-en-7b is fine-tuned based on Qwen1.5-7B with English instruction data, and firefly-qwen1.5-en-7b-dpo-v0.1 is trained with [Direct Preference Optimization (DPO)](https://arxiv.org/abs/2305.18290) based on firefly-qwen1.5-en-7b.
29
+
30
+ 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).
31
+
32
+ <img src="pics/open_llm.png" width="800">
33
+
34
+ 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
35
+ the performance in Chinese yet.
36
+
37
+ We advise you to install transformers>=4.37.0.
38
+
39
+ ## Performance
40
+ We evaluate our models on [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard), they achieve good performance.
41
+
42
+ | Model | Average | ARC | HellaSwag | MMLU | TruthfulQA | Winogrande | GSM8K |
43
+ |-----------------------------------|--------|--------|-----------|--------|------------|------------|--------|
44
+ | firefly-gemma-7b | 62.93 | 62.12 | 79.77 | 61.57 | 49.41 | 75.45 | 49.28 |
45
+ | **firefly-qwen1.5-en-7b-dpo-v0.1** | 62.36 | 54.35 | 76.04 | 61.21 | 56.4 | 72.06 | 54.13 |
46
+ | zephyr-7b-beta | 61.95 | 62.03 | 84.36 | 61.07 | 57.45 | 77.74 | 29.04 |
47
+ | **firefly-qwen1.5-en-7b** | 61.44 | 53.41 | 75.51 | 61.67 |51.96 |70.72 | 55.34 |
48
+ | vicuna-13b-v1.5 | 55.41 | 57.08 | 81.24 | 56.67 | 51.51 | 74.66 | 11.3 |
49
+ | Xwin-LM-13B-V0.1 | 55.29 | 62.54 | 82.8 | 56.53 | 45.96 | 74.27 | 9.63 |
50
+ | Qwen1.5-7B-Chat | 55.15 | 55.89 | 78.56 | 61.65 | 53.54 | 67.72 | 13.57 |
51
+ | gemma-7b-it | 53.56 | 51.45 | 71.96 | 53.52 | 47.29 | 67.96 | 29.19 |
52
+
53
+
54
+
55
+ ## Usage
56
+ The chat templates of our chat models are the same as Official Qwen1.5-7B-Chat:
57
+ ```text
58
+ <|im_start|>system
59
+ You are a helpful assistant.<|im_end|>
60
+ <|im_start|>user
61
+ hello, who are you?<|im_end|>
62
+ <|im_start|>assistant
63
+ I am a AI program developed by Firefly<|im_end|>
64
+ ```
65
+
66
+ You can use script to inference in [Firefly](https://github.com/yangjianxin1/Firefly/blob/master/script/chat/chat.py).
67
+
68
+ You can also use the following code:
69
+ ```python
70
+ from transformers import AutoModelForCausalLM, AutoTokenizer
71
+ import torch
72
+
73
+ model_name_or_path = "YeungNLP/firefly-qwen1.5-en-7b"
74
+ model = AutoModelForCausalLM.from_pretrained(
75
+ model_name_or_path,
76
+ trust_remote_code=True,
77
+ low_cpu_mem_usage=True,
78
+ torch_dtype=torch.float16,
79
+ device_map='auto',
80
+ )
81
+ tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
82
+
83
+ prompt = "Compose an engaging travel blog post about a recent trip to Hawaii, highlighting cultural experiences and must-see attractions. "
84
+ messages = [
85
+ {"role": "system", "content": "You are a helpful assistant."},
86
+ {"role": "user", "content": prompt}
87
+ ]
88
+ text = tokenizer.apply_chat_template(
89
+ messages,
90
+ tokenize=False,
91
+ add_generation_prompt=True
92
+ )
93
+ model_inputs = tokenizer([text], return_tensors="pt").to('cuda')
94
+
95
+ generated_ids = model.generate(
96
+ model_inputs.input_ids,
97
+ max_new_tokens=1500,
98
+ top_p = 0.9,
99
+ temperature = 0.35,
100
+ repetition_penalty = 1.0,
101
+ eos_token_id=tokenizer.encode('<|im_end|>', add_special_tokens=False)
102
+ )
103
+ generated_ids = [
104
+ output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
105
+ ]
106
+
107
+ response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
108
+ print(response)
109
+ ```
110
+
111
+ ## Training Details
112
+ Both in SFT and DPO stages, **We only use a single V100 GPU** with QLoRA, and we use [Firefly](https://github.com/yangjianxin1/Firefly) to train our models.
113
+
114
+ ### Training Setting
115
+ The following hyperparameters are used during SFT:
116
+ - num_epochs: 1
117
+ - learning_rate: 2e-4
118
+ - total_train_batch_size: 32
119
+ - max_seq_length: 2048
120
+ - optimizer: paged_adamw_32bit
121
+ - lr_scheduler_type: constant_with_warmup
122
+ - warmup_steps: 700
123
+ - lora_rank: 64
124
+ - lora_alpha: 16
125
+ - lora_dropout: 0.05
126
+ - gradient_checkpointing: true
127
+ - fp16: true
128
+
129
+ The following hyperparameters were used during DPO:
130
+ - num_epochs: 1
131
+ - learning_rate: 2e-4
132
+ - total_train_batch_size: 32
133
+ - max_seq_length: 1600
134
+ - max_prompt_length: 500
135
+ - optimizer: paged_adamw_32bit
136
+ - lr_scheduler_type: constant_with_warmup
137
+ - warmup_steps: 200
138
+ - lora_rank: 64
139
+ - lora_alpha: 16
140
+ - lora_dropout: 0.05
141
+ - gradient_checkpointing: true
142
+ - fp16: true
143
+
144
+
145
+ ### Training metrics
146
+ Training Rewards/margins in DPO:
147
+
148
+ <img src="pics/margins.png" width="600">
149
+
150
+ Training Rewards/accuracies in DPO:
151
+
152
+ <img src="pics/accuracies.png" width="500">
153
+
154
+ Training loss in DPO:
155
+
156
+ <img src="pics/loss.png" width="500">
157
+
158
+ The table below shows the full set of DPO training metrics:
159
+
160
+ | Epoch | Step | Loss | Rewards/accuracies | Rewards/margins | Rewards/chosen | Rewards/rejected | Logits/chosen| Logits/rejected | Logps/chosen| Logps/rejected|
161
+ |---|---|---|---|---|---|---|---|---|---|---|
162
+ |0.05|100|0.6231|0.6587|0.3179|0.0404|-0.2774|1.1694|1.2377|-284.5586|-255.4863|
163
+ |0.1|200|0.5945|0.6894|0.5988|-0.1704|-0.7693|1.012|1.0283|-284.3049|-268.1887|
164
+ |0.16|300|0.5754|0.6981|0.8314|-0.282|-1.1133|0.8912|0.8956|-283.6926|-270.3117|
165
+ |0.21|400|0.5702|0.7194|0.9369|-0.1944|-1.1313|0.7255|0.7557|-291.2833|-273.9706|
166
+ |0.26|500|0.5913|0.695|0.8784|-0.4524|-1.3309|0.5491|0.5535|-289.5705|-271.754|
167
+ |0.31|600|0.5743|0.6994|1.0192|-0.4505|-1.4698|0.6446|0.6399|-296.5292|-277.824|
168
+ |0.37|700|0.5876|0.7219|1.0471|-0.6998|-1.747|0.4955|0.4329|-303.7684|-289.0117|
169
+ |0.42|800|0.5831|0.715|1.0485|-0.8185|-1.8671|0.5589|0.4804|-295.6313|-288.0656|
170
+ |0.47|900|0.5674|0.7119|1.1854|-1.2085|-2.3939|0.3467|0.2249|-302.3643|-286.2816|
171
+ |0.52|1000|0.5794|0.7138|1.1458|-0.8423|-1.9881|0.5116|0.4248|-299.3136|-287.3934|
172
+ |0.58|1100|0.5718|0.7194|1.2897|-1.4944|-2.7841|0.6392|0.5739|-316.6829|-294.1148|
173
+ |0.63|1200|0.5718|0.7275|1.2459|-1.7543|-3.0002|0.4999|0.4065|-316.7873|-297.8514|
174
+ |0.68|1300|0.5789|0.72|1.3379|-1.8485|-3.1864|0.4289|0.3172|-314.8326|-296.8319|
175
+ |0.73|1400|0.5462|0.7425|1.4074|-1.9865|-3.3939|0.3645|0.2333|-309.4503|-294.3931|
176
+ |0.79|1500|0.5829|0.7156|1.2582|-2.1183|-3.3766|0.4193|0.2796|-307.5281|-292.0817|
177
+ |0.84|1600|0.5575|0.7375|1.471|-2.1429|-3.6139|0.6547|0.5152|-310.9912|-298.899|
178
+ |0.89|1700|0.5638|0.745|1.5433|-2.991|-4.5343|0.7336|0.6782|-328.2657|-307.5182|
179
+ |0.94|1800|0.5559|0.7181|1.4484|-2.8818|-4.3302|0.7997|0.8327|-316.2716|-295.1836|
180
+ |0.99|1900|0.5627|0.7387|1.5378|-2.7941|-4.332|0.8573|0.858|-324.9405|-310.1192|
181
+